A model is still running

A model is still running or downloading. Leaving will stop the current operation. Stay on this page to let it finish.

Skip to content
LocalAI.toolsBeta

Join the waitlist

Get an email when the new apps and model runs launch.

No account needed.

Canary 2

Align a matching transcript to its recording with the auxiliary CTC checkpoint distributed with NVIDIA Canary 2.

Desktop inferencePrivate · Runs on your computer
About this model

Align a matching transcript to its recording with the auxiliary CTC checkpoint distributed with NVIDIA Canary 2.

Created by
NVIDIA
License
CC-BY-4.0
Model released
Aug 8, 2025
Source repository created
Aug 4, 2025
Model weights · Q4
392.2 MB

Best for

  • Correct subtitle word timing against an existing transcript
  • Inspect local transcript/audio alignment with human review

Keep in mind

  • Requires an accurate transcript in the same language as the recording. Translated text cannot be aligned to different-language speech.
  • Up to 30 seconds of mono 16 kHz audio and 4000 characters per request. Whitespace-separated words must fit 63 UTF-8 bytes.
  • 80 ms encoder frames limit timestamp resolution. Review boundaries before overwriting subtitles.
  • English single-recording clean/noise/shift pilot only. Other languages and recordings remain unverified.
  • Silence and incorrect transcripts can produce plausible incorrect boundaries. Not suitable for forensic or high-impact decisions.
  • Metal uses CPU preprocessing, Viterbi alignment and fallback operations; acceleration is partial.
  • The runtime computes the full CTC path before returning word timings; incremental output is unavailable.

Processed on your deviceYour input is processed locally in the desktop app. Model and runtime files download from external hosts.

Target platforms · verification required

  • macos · aarch64, x86_64
  • windows · x86_64

Recommended configuration

Q4 · metal (verified platform only)

Sources & technical details
Original model
https://huggingface.co/nvidia/canary-1b-v2
Release verification
Verified
Release evidence
Auxiliary timestamps_asr_model_weights.ckpt distributed inside the official Canary 1B v2 NeMo archive; this is not the main encoder-decoder ASR checkpoint. See NVIDIA technical report section 5.
Source updated
Aug 31, 2026
Native model files
cstr/canary-ctc-aligner-GGUF
Files updated
Aug 2, 2026
Artifact revision
2b22dc9aff585bc8368a5228173c8afe22c155b7

Desktop app

Sample result

Canary 2 CTC Aligner 0.6B

Run Canary 2 on your computer

Download the desktop app to load this model, provide input and view results locally.

Public desktop downloads are coming soon.

Device requirements

macOS Apple SiliconCPU / Q4Tested8.0 GiBRAM estimate
macOS Apple SiliconMETAL / Q4Tested8.0 GiBRAM estimate
Windows Intel / AMDCPU / Q4Unverified8.0 GiBRAM estimate
macOS Intel / AMDCPU / Q4Unverified8.0 GiBRAM estimate
If your device falls short

MemoryAdvisory only: close other apps, reduce image/tile/token/batch settings or choose a smaller verified model/precision. CPU can need more RAM than GPU.

GPUIf the selected GPU runtime fails, existing CPU fallback may recover when enabled, but CPU has its own memory/runtime needs.

PlatformUse a compatible OS/architecture/runtime or another model. These estimates do not add hardware detection or block runs.

Benchmark evidence

Saved desktop measurements apply to the recorded hardware and runtime. Fallback cases report the device that actually ran; they do not prove GPU support on the requested device.

Canary 2 CTC Aligner 0.6B

72 configurations tested
Passed72
Failed0
TestedNov 3, 2026
Runtimecrispasr 0.8.50

Quality criteria

{ "scope": "Initial English forced-alignment screening, frozen before inference; no broad multilingual claim", "gates": { "exactTranscriptTokenCoverage": true, "allTimesFinite": true, "nonnegativeMonotonicSpans": true, "endWithinAudioToleranceSeconds": 0.24, "maximumZeroDurationFraction": 0.1, "maximumReferenceBoundaryMAESeconds": 0.2, "maximumRealTimeFactor": 1 }, "negativeCases": [ "empty transcript rejected", "nonfinite audio rejected", "unknown configuration rejected", "words over 63 UTF-8 bytes rejected" ], "performancePolicy": "At least realtime warm alignment on a short clean English recording. Reference boundary quality requires an independent saved reference before support can be advertised." }

PrecisionDeviceResultPlatformInference / loadTask metric
Q4metalAcceleratedPassedVery fastmacos · aarch64111 ms / 619 ms0.010 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64113 ms / 619 ms0.010 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64114 ms / 608 ms0.010 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64114 ms / 608 ms0.010 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64115 ms / 619 ms0.010 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64115 ms / 619 ms0.010 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64115 ms / 619 ms0.010 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64116 ms / 619 ms0.011 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64116 ms / 629 ms0.011 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64117 ms / 629 ms0.011 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64117 ms / 629 ms0.011 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64124 ms / 619 ms0.011 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64119 ms / 608 ms0.011 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64119 ms / 608 ms0.011 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64119 ms / 629 ms0.011 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64120 ms / 629 ms0.011 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64120 ms / 608 ms0.011 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64128 ms / 608 ms0.011 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64126 ms / 629 ms0.011 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64127 ms / 608 ms0.012 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64129 ms / 629 ms0.012 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64134 ms / 608 ms0.012 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64136 ms / 619 ms0.012 RTF
Q4metalAcceleratedPassedVery fastmacos · aarch64267 ms / 629 ms0.023 RTF
Q4cpuPassedVery fastmacos · aarch64583 ms / 575 ms0.053 RTF
Q4cpuPassedVery fastmacos · aarch64584 ms / 528 ms0.053 RTF
Q4cpuPassedVery fastmacos · aarch64585 ms / 541 ms0.053 RTF
Q4cpuPassedVery fastmacos · aarch64585 ms / 528 ms0.053 RTF
Q4cpuPassedVery fastmacos · aarch64585 ms / 575 ms0.053 RTF
Q4cpuPassedVery fastmacos · aarch64586 ms / 528 ms0.053 RTF
Q4cpuPassedVery fastmacos · aarch64586 ms / 541 ms0.053 RTF
Q4cpuPassedVery fastmacos · aarch64586 ms / 528 ms0.053 RTF
Q4cpuPassedVery fastmacos · aarch64586 ms / 575 ms0.053 RTF
Q4cpuPassedVery fastmacos · aarch64587 ms / 528 ms0.053 RTF
Q4cpuPassedVery fastmacos · aarch64587 ms / 575 ms0.053 RTF
Q4cpuPassedVery fastmacos · aarch64588 ms / 575 ms0.053 RTF
Q4cpuPassedVery fastmacos · aarch64589 ms / 541 ms0.054 RTF
Q4cpuPassedVery fastmacos · aarch64589 ms / 541 ms0.054 RTF
Q4cpuPassedVery fastmacos · aarch64590 ms / 541 ms0.054 RTF
Q4cpuPassedVery fastmacos · aarch64590 ms / 528 ms0.054 RTF
Q4cpuPassedVery fastmacos · aarch64591 ms / 541 ms0.054 RTF
Q4cpuPassedVery fastmacos · aarch64620 ms / 541 ms0.054 RTF
Q4cpuPassedVery fastmacos · aarch64597 ms / 528 ms0.054 RTF
Q4cpuPassedVery fastmacos · aarch64600 ms / 541 ms0.055 RTF
Q4cpuPassedVery fastmacos · aarch64628 ms / 528 ms0.055 RTF
Q4cpuPassedVery fastmacos · aarch64603 ms / 575 ms0.055 RTF
Q4cpuPassedVery fastmacos · aarch64627 ms / 575 ms0.057 RTF
Q4cpuPassedVery fastmacos · aarch64677 ms / 575 ms0.059 RTF
Q4cpuPassedFastmacos · aarch642033 ms / 10888 ms0.185 RTF
Q4cpuPassedFastmacos · aarch642045 ms / 10888 ms0.186 RTF
Q4cpuPassedFastmacos · aarch642046 ms / 531 ms0.186 RTF
Q4cpuPassedFastmacos · aarch642051 ms / 10888 ms0.186 RTF
Q4cpuPassedFastmacos · aarch642052 ms / 531 ms0.187 RTF
Q4cpuPassedFastmacos · aarch642057 ms / 536 ms0.187 RTF
Q4cpuPassedFastmacos · aarch642151 ms / 536 ms0.187 RTF
Q4cpuPassedFastmacos · aarch642164 ms / 10888 ms0.188 RTF
Q4cpuPassedFastmacos · aarch642075 ms / 531 ms0.189 RTF
Q4cpuPassedFastmacos · aarch642084 ms / 10888 ms0.189 RTF
Q4cpuPassedFastmacos · aarch642085 ms / 536 ms0.190 RTF
Q4cpuPassedFastmacos · aarch642086 ms / 536 ms0.190 RTF
Q4cpuPassedFastmacos · aarch642086 ms / 531 ms0.190 RTF
Q4cpuPassedFastmacos · aarch642086 ms / 531 ms0.190 RTF
Q4cpuPassedFastmacos · aarch642181 ms / 531 ms0.190 RTF
Q4cpuPassedFastmacos · aarch642086 ms / 536 ms0.190 RTF
Q4cpuPassedFastmacos · aarch642087 ms / 10888 ms0.190 RTF
Q4cpuPassedFastmacos · aarch642087 ms / 536 ms0.190 RTF
Q4cpuPassedFastmacos · aarch642091 ms / 531 ms0.190 RTF
Q4cpuPassedFastmacos · aarch642092 ms / 10888 ms0.190 RTF
Q4cpuPassedFastmacos · aarch642098 ms / 10888 ms0.191 RTF
Q4cpuPassedFastmacos · aarch642118 ms / 536 ms0.193 RTF
Q4cpuPassedFastmacos · aarch642130 ms / 536 ms0.194 RTF
Q4cpuPassedFastmacos · aarch642153 ms / 531 ms0.196 RTF
Test setup & sample details
Quality gate
{ "scope": "Initial English forced-alignment screening, frozen before inference; no broad multilingual claim", "gates": { "exactTranscriptTokenCoverage": true, "allTimesFinite": true, "nonnegativeMonotonicSpans": true, "endWithinAudioToleranceSeconds": 0.24, "maximumZeroDurationFraction": 0.1, "maximumReferenceBoundaryMAESeconds": 0.2, "maximumRealTimeFactor": 1 }, "negativeCases": [ "empty transcript rejected", "nonfinite audio rejected", "unknown configuration rejected", "words over 63 UTF-8 bytes rejected" ], "performancePolicy": "At least realtime warm alignment on a short clean English recording. Reference boundary quality requires an independent saved reference before support can be advertised." }
Tested
2026-11-03T18:47:33.301Z
Environment
{ "hardware": "Apple M3 Max", "os": "27.2", "architecture": "aarch64", "runtime": "CrispASR 0.8.50 Q4_K", "interpreterScope": "App-owned isolated CPython 3.13.18; public DesktopAI model-host; debug build" }
Runtime
crispasr 0.8.50
Artifact revision
2b22dc9aff585bc8368a5228173c8afe22c155b7
Fixture
{
  "file": "model-samples/desktop/qwen3-forced-aligner/clean.wav",
  "url": "https://assets.localai.tools/samples/v1/files/ba1a561773d6b6e55c7b9d326d9aa792f6d3a6f430556b83a78f78a2a4cb25e1/clean.wav",
  "size": 352044,
  "sha1": "ce87ae892606eecb8ef65803ed7c4d1ea13ca437",
  "sha256": "ba1a561773d6b6e55c7b9d326d9aa792f6d3a6f430556b83a78f78a2a4cb25e1"
}
Rerun when
Exact artifact/runtime revision; Audio/transcript/reference checksum or quality gates; Task preprocessing/decoding or backend behavior
Per-case quality & failure details
[
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2166.35825,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2097.948874987196,
      "loadMs": 10887.648000000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.190722624998836,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 12985.596874987197,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2118.8029579999998,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2051.457916997606,
      "loadMs": 10887.648000000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.18649617427250964,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 12939.105916997607,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2111.678041,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2044.9987500032876,
      "loadMs": 10887.648000000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.18590897727302613,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 12932.646750003289,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2159.019958,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2092.086874996312,
      "loadMs": 10887.648000000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.19018971590875564,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 12979.734874996313,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2154.840917,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2086.7260829836596,
      "loadMs": 10887.648000000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.1897023711803327,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 12974.37408298366,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2151.562083,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2083.7513750011567,
      "loadMs": 10887.648000000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.18943194318192333,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 12971.399375001158,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2098.736916,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2032.5939579925032,
      "loadMs": 10887.648000000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.1847812689084094,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 12920.241957992504,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.0940909090909091,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11.5,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2231.114916,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2164.102292008465,
      "loadMs": 10887.648000000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.1881828080007361,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 13051.750292008466,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2124.363292,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2056.7370420030784,
      "loadMs": 535.630625,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.18697609472755256,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2592.367667003078,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2152.21725,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2085.1886669988744,
      "loadMs": 535.630625,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.18956260609080677,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2620.819291998874,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2154.67225,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2085.640749981394,
      "loadMs": 535.630625,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.1896037045437631,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2621.271374981394,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2154.834917,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2086.417457991047,
      "loadMs": 535.630625,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.18967431436282245,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2622.0480829910466,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2156.0205419999998,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2087.4962499947287,
      "loadMs": 535.630625,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.18977238636315716,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2623.1268749947285,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2196.0476249999997,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2129.651583993109,
      "loadMs": 535.630625,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.19360468945391895,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2665.2822089931087,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2188.5761669999997,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2118.452624999918,
      "loadMs": 535.630625,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.1925866022727198,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2654.083249999918,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.0940909090909091,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11.5,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2219.042542,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2150.8013749844395,
      "loadMs": 535.630625,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.18702620652038604,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2686.4319999844392,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2221.115208,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2152.865959011251,
      "loadMs": 531.227583,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.195715087182841,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2684.093542011251,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2154.621291,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2086.1682499817107,
      "loadMs": 531.227583,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.1896516590892464,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2617.3958329817106,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2160.0513339999998,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2091.40070801368,
      "loadMs": 531.227583,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.1901273370921527,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2622.6282910136797,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2113.75125,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2046.124583983328,
      "loadMs": 531.227583,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.1860113258166662,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2577.352166983328,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2118.235125,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2051.560874999268,
      "loadMs": 531.227583,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.18650553409084256,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2582.788457999268,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2153.676291,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2086.061707988847,
      "loadMs": 531.227583,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.1896419734535315,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2617.289290988847,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2142.1713750000004,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2074.779458023841,
      "loadMs": 531.227583,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.1886163143658037,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2606.007041023841,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.0940909090909091,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11.5,
      "cpuThreads": 1,
      "device": "cpu",
      "durationMs": 2250.1146670000003,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 2181.1782080039848,
      "loadMs": 531.227583,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.1896676702612161,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 2712.4057910039846,
      "grade": "good"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 670.082833,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 602.5336250022519,
      "loadMs": 575.4275,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.05477578409111381,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1177.961125002252,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 700.3039170000001,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 627.1862080029678,
      "loadMs": 575.4275,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.0570169280002698,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1202.6137080029678,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 652.7315,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 585.2432079846039,
      "loadMs": 575.4275,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.05320392799860036,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1160.670707984604,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 654.71225,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 588.1631249794737,
      "loadMs": 575.4275,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.05346937499813397,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1163.5906249794737,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 652.1587089999999,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 586.2841669877525,
      "loadMs": 575.4275,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.05329856063525023,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1161.7116669877526,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 652.6831669999999,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 586.6727079846896,
      "loadMs": 575.4275,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.05333388254406269,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1162.1002079846896,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 650.664333,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 583.254291996127,
      "loadMs": 575.4275,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.053023117454193365,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1158.681791996127,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.0940909090909091,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11.5,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 743.452041,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 676.6570000036154,
      "loadMs": 575.4275,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.05883973913074917,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1252.0845000036154,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 666.501875,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 600.0687920022756,
      "loadMs": 540.5534170000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.05455170836384324,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1140.6222090022757,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 658.578333,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 589.5757499965839,
      "loadMs": 540.5534170000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.0535977954542349,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1130.129166996584,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 660.150583,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 589.3691669916734,
      "loadMs": 540.5534170000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.05357901518106122,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1129.9225839916735,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 654.480792,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 585.8828340133186,
      "loadMs": 540.5534170000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.0532620758193926,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1126.4362510133187,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 658.286209,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 591.3652500021271,
      "loadMs": 540.5534170000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.05376047727292065,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1131.9186670021272,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 657.479625,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 589.1410839976743,
      "loadMs": 540.5534170000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.053558280363424936,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1129.6945009976744,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 655.301875,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 584.776416013483,
      "loadMs": 540.5534170000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.05316149236486209,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1125.329833013483,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.0940909090909091,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11.5,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 688.6387090000001,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 620.044000010239,
      "loadMs": 540.5534170000001,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.053916869566107736,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1160.597417010239,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 665.1509159999999,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 597.3600409924984,
      "loadMs": 527.982666,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.05430545827204531,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1125.3427069924983,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 654.482958,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 585.7156660058536,
      "loadMs": 527.982666,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.05324687872780487,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1113.6983320058534,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 652.663709,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 585.0085830024909,
      "loadMs": 527.982666,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.053182598454771905,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1112.9912490024908,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 651.253083,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 583.7479170004372,
      "loadMs": 527.982666,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.0530679924545852,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1111.730583000437,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 657.578416,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 589.7541249869391,
      "loadMs": 527.982666,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.05361401136244901,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1117.736790986939,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.08954545454545454,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 652.584125,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 586.6222080076113,
      "loadMs": 527.982666,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.053329291637055576,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1114.6048740076112,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 652.3915000000001,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 586.0306250106078,
      "loadMs": 527.982666,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.05327551136460071,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1114.0132910106076,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "cpu",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.0940909090909091,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "cpu",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11.5,
      "cpuThreads": 4,
      "device": "cpu",
      "durationMs": 695.3965840000001,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 628.1479579920415,
      "loadMs": 527.982666,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "crispasr_init_gpu_backend: --gpu-backend cpu — using the CPU backend, no GPU device initialised\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "CPU",
      "realTimeFactor": 0.05462156156452535,
      "requestedDevice": "cpu",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 1156.1306239920414,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 197.131292,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 128.86683299439028,
      "loadMs": 629.313209,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.016 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.011715166635853662,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 758.1800419943903,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 189.970625,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 119.31104201357812,
      "loadMs": 629.313209,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.016 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.010846458364870738,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 748.6242510135781,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 193.687,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 125.92449999647216,
      "loadMs": 629.313209,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.016 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.011447681817861105,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 755.2377089964722,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 184.34375,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 117.0055829861667,
      "loadMs": 629.313209,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.016 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.010636871180560607,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 746.3187919861667,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 184.24375,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 116.23858302482404,
      "loadMs": 629.313209,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.016 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.01056714391134764,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 745.5517920248241,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 184.804208,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 117.2875830088742,
      "loadMs": 629.313209,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.016 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.010662507546261291,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 746.6007920088742,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 187.58783400000002,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 119.5604579988867,
      "loadMs": 629.313209,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.016 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.010869132545353336,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 748.8736669988867,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.0940909090909091,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11.5,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 335.17516600000005,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 266.77583300624974,
      "loadMs": 629.313209,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.016 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.016 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.023197898522282587,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 896.0890420062498,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 201.481875,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 133.78454200574197,
      "loadMs": 608.231709,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.015 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.012162231091431087,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 742.016251005742,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 187.987042,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 119.0167089807801,
      "loadMs": 608.231709,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.015 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.010819700816434552,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 727.2484179807801,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 189.181625,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 120.44358300045133,
      "loadMs": 608.231709,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.015 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.010949416636404669,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 728.6752920004513,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 195.44262500000002,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 127.23166699288412,
      "loadMs": 608.231709,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.015 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.011566515181171284,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 735.4633759928842,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 181.827875,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 113.88245798298158,
      "loadMs": 608.231709,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.015 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.010352950725725596,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 722.1141669829816,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 181.59545899999998,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 114.487916987855,
      "loadMs": 608.231709,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.015 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.010407992453441362,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 722.719625987855,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 186.5025,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 119.13725000340492,
      "loadMs": 608.231709,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.015 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.01083065909121863,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 727.3689590034049,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.0940909090909091,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11.5,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 198.33166699999998,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 127.69374999334104,
      "loadMs": 608.231709,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.015 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.010 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.004 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.015 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.011103804347247047,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 735.9254589933411,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 202.882375,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 135.64412502455525,
      "loadMs": 618.9075,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.003 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.014 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.012331284093141388,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 754.5516250245553,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 184.136333,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 115.018999989843,
      "loadMs": 618.9075,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.003 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.014 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.010456272726349363,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 733.926499989843,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 183.103959,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 114.64250000426546,
      "loadMs": 618.9075,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.003 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.014 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.010422045454933224,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 733.5500000042655,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 181.37975,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 112.9323749919422,
      "loadMs": 618.9075,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.003 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.014 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.010266579544722015,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 731.8398749919422,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 182.44433400000003,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 115.6377920124214,
      "loadMs": 618.9075,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.003 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.014 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.010512526546583766,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 734.5452920124214,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 179.0815,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 111.4845419942867,
      "loadMs": 618.9075,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.003 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.014 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.010134958363116972,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 730.3920419942867,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.09136363636363636,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 182.73875,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 115.35804098821244,
      "loadMs": 618.9075,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.003 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.014 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.010487094635292042,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 734.2655409882125,
      "grade": "excellent"
    }
  },
  {
    "precision": "q4",
    "device": "metal",
    "status": "pass",
    "quality": {
      "passed": true,
      "boundaryMAESeconds": 0.0940909090909091,
      "referenceType": "ElevenLabsScribeV2Agreement"
    },
    "runtime": {
      "acceleration": "partial",
      "accelerationDetails": "CTC encoder uses selected ggml backend; unsupported ops, audio preprocessing and Viterbi alignment use CPU.",
      "artifactRevision": "2b22dc9aff585bc8368a5228173c8afe22c155b7",
      "audioSeconds": 11.5,
      "cpuThreads": 4,
      "device": "metal",
      "durationMs": 192.540917,
      "engine": "crispasr",
      "fallback": false,
      "fallbackReason": null,
      "family": {
        "catalogId": "canary-v2-ctc-alignment",
        "defaultModelId": "canary-2-ctc-aligner-0.6b-native",
        "displayName": "Canary 2 CTC Aligner",
        "id": "canary-v2",
        "variantId": "0-6b-ctc",
        "variantName": "0.6B · CTC"
      },
      "inferenceMs": 124.21820900635794,
      "loadMs": 618.9075,
      "modelId": "canary-2-ctc-aligner-0.6b-native",
      "placement": "ggml_metal_device_init: tensor API disabled for pre-M5 and pre-A19 devices\ncrispasr_metal_pipeline_cache_open: pipeline cache loaded: /Users/pengzhang/Library/Caches/ggml-metal/Apple_M3_Max.archive\nggml_metal_library_init: using embedded metal library\nggml_metal_library_compile_all: compiled 'fa_aux' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_f16' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q4_0' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_q5_0' library in 0.012 sec\nggml_metal_library_compile_all: compiled 'fa_q5_1' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_q8_0' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f16' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'fa_vec_f32' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_0' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q4_1' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_0' library in 0.014 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q5_1' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'fa_vec_q8_0' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'mul_mv' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'mul_mv_mma' library in 0.013 sec\nggml_metal_library_compile_all: compiled 'mul_mm' library in 0.005 sec\nggml_metal_library_compile_all: compiled 'quantize' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'softmax' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'norm' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'unary' library in 0.002 sec\nggml_metal_library_compile_all: compiled 'binbcast' library in 0.011 sec\nggml_metal_library_compile_all: compiled 'reduce' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'tri' library in 0.007 sec\nggml_metal_library_compile_all: compiled 'ssm' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'wkv' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'gated_delta_net' library in 0.009 sec\nggml_metal_library_compile_all: compiled 'solve_tri' library in 0.000 sec\nggml_metal_library_compile_all: compiled 'rope' library in 0.008 sec\nggml_metal_library_compile_all: compiled 'conv' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'aa_snake_beta' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'upscale' library in 0.003 sec\nggml_metal_library_compile_all: compiled 'argsort' library in 0.004 sec\nggml_metal_library_compile_all: compiled 'pool' library in 0.006 sec\nggml_metal_library_compile_all: compiled 'misc' library in 0.003 sec\nggml_metal_library_compile_all: loaded 36 libraries from embedded data in 0.016 sec (max single = 0.014 sec)\nggml_metal_rsets_init: creating a residency set collection (keep_alive = 180 s)\nggml_metal_device_init: GPU name:   MTL0 (Apple M3 Max)\nggml_metal_device_init: GPU family: MTLGPUFamilyApple9  (1009)\nggml_metal_device_init: GPU family: MTLGPUFamilyMetal4  (5002)\nggml_metal_device_init: simdgroup reduction   = true\nggml_metal_device_init: simdgroup matrix mul. = true\nggml_metal_device_init: has unified memory    = true\nggml_metal_device_init: has bfloat            = true\nggml_metal_device_init: has tensor            = false\nggml_metal_device_init: use residency sets    = true\nggml_metal_device_init: use shared buffers    = true\nggml_metal_device_init: recommendedMaxWorkingSetSize  = 83494.17 MB\nggml_metal_init: allocating\nggml_metal_init: found device: Apple M3 Max\nggml_metal_init: picking default device: Apple M3 Max\nggml_metal_init: use fusion         = true\nggml_metal_init: use concurrency    = true\nggml_metal_init: use graph optimize = true\ncrispasr_init_gpu_backend: using preferred GPU backend: MTL0\nggml_metal_log_allocated_size: allocated buffer, size =   374.00 MiB, (  374.45 / 79626.25)\ncanary_ctc: vocab=16384  d_model=1024  n_layers=24  n_heads=8  ff=4096\ncanary_ctc: BN folded into conv_dw weights for 24 layers\ncanary_ctc: fused Q/K/V projections for 24 layers (CRISPASR_FC_FUSED_QKV)\n",
      "precision": "q4",
      "provider": "Metal",
      "realTimeFactor": 0.010801583391857212,
      "requestedDevice": "metal",
      "requestedPrecision": "q4",
      "reused": false,
      "sampleRate": 16000,
      "selectionReason": "Explicit device selection; actual placement is reported after inference",
      "version": "0.8.50",
      "totalMs": 743.125709006358,
      "grade": "excellent"
    }
  }
]

Something not quite right?

Tell us what happened with Canary 2. Loading, results, or the page itself—we’re here for all of it.

Attachments Optional

Up to 3 files · 50 MB each

Browser, OS, CPU/GPU, memory, screen, storage, and network details are included automatically. Model inputs, results, and browsing history are not collected.