{
  "measurement": "On-device transcription throughput, expressed as a multiple of realtime (higher is faster: 12.4x means one hour of audio is processed in about 4 minutes 50 seconds of work)",
  "published_on": "2026-09-18",
  "app_repository": "https://github.com/ComicBit/Aurora",
  "rows": [
    {
      "measurement": "end_to_end_offline_refinement",
      "device": "iPhone 15 Pro",
      "compute_unit": "Neural Engine",
      "recognition_model": "distil-whisper_distil-large-v3_turbo_600MB (tier distilled-v3)",
      "diarization_model": "community-1 offline diarizer",
      "audio_seconds": 97.432,
      "decode_seconds": 7.673,
      "diarize_seconds": 5.223,
      "wall_seconds": 7.858,
      "realtime_factor": 12.4,
      "measured_on": "2026-08-26",
      "run": "fixture run, app launch 5301, device 'localghost'",
      "evidence": "docs/plans/unified-refinement-pipeline.md:299-312 (Aurora app repository)"
    },
    {
      "measurement": "whisper_decode_backend_comparison_neural_engine",
      "device": "Apple silicon Mac (M2 Pro)",
      "compute_unit": "Neural Engine",
      "recognition_model": "WhisperKit, .vad decoding strategy",
      "diarization_model": "",
      "audio_seconds": "",
      "decode_seconds": "",
      "diarize_seconds": "",
      "wall_seconds": "",
      "realtime_factor": 22.0,
      "measured_on": "2026-08-26",
      "run": "decode-path comparison for a GPU-placement decision; NOT end-to-end refinement",
      "evidence": "docs/plans/unified-refinement-pipeline.md:512-513 (Aurora app repository)"
    },
    {
      "measurement": "whisper_decode_backend_comparison_gpu_mlx",
      "device": "Apple silicon Mac (M2 Pro)",
      "compute_unit": "GPU via MLX",
      "recognition_model": "MLX sequential Whisper",
      "diarization_model": "",
      "audio_seconds": "",
      "decode_seconds": "",
      "diarize_seconds": "",
      "wall_seconds": "",
      "realtime_factor": 11.2,
      "measured_on": "2026-08-26",
      "run": "decode-path comparison for a GPU-placement decision; NOT end-to-end refinement",
      "evidence": "docs/plans/unified-refinement-pipeline.md:512-513 (Aurora app repository)"
    }
  ],
  "not_measured": [
    "Live transcription throughput. The live path must sustain better than 1x realtime to keep up with speech, and in practice it does, but no published measurement of its sustained factor exists. The site previously printed '> 1x sustained' as a measured figure; it was not one.",
    "Offline refinement end-to-end on an Apple silicon Mac. The 22x and 11.2x rows above compare Whisper decode backends on an M2 Pro for a GPU-placement decision; they are not an end-to-end refinement measurement, and the Mac app is still in development.",
    "Offline refinement on iPhone models other than the iPhone 15 Pro.",
    "Peak memory of the app during recording or refinement.",
    "Battery drain per hour of recording or refinement."
  ],
  "limitations": [
    "One fixture, one device, one run. 97 seconds of audio is a short fixture for a throughput number.",
    "Throughput is model-dependent: the 12.4x row used the 600 MB distil-large-v3-turbo recognizer, which is the largest tier, not the onboarding default.",
    "Thermal state and Low Power Mode were not recorded for the run and both change the result.",
    "Provisional: no independent verification pass has run."
  ]
}