{
  "runtime": "e801e7d6c2692121a39e80ae525cb5265174a495",
  "exports": {
    "stephvax": "fc68ee7e5a0a29662e218df84b24acb311f7fb6d",
    "kevin": "58a6d00cf13d239b6748cb0769f35c580a8f606c"
  },
  "cpuAffinity": [
    0,
    2
  ],
  "inferenceThreads": 2,
  "temperature": 0.3,
  "lsdSteps": 1,
  "cppOrt": "1.23.2",
  "kevinOrt": "1.24.4",
  "notes": [
    "C++ uses benchmark compatibility patch with BOS appended inside encoder graph, mixed precision, short-input space padding disabled, disk cache disabled, seed and chunk-size controls.",
    "English bundle recommends no fixed extra EOS override; existing C++ 5/3 extra frames match Kevin guess+2.",
    "C++ uses its sentence splitter; it does not enforce the bundle 50-token limit. Holdout includes a long sentence.",
    "C++ language weights INT8; mixed flow/decoder FP32; full INT8 is a separate comparison.",
    "Onset filtering -50 dBFS 10ms window, 40ms preroll, 2s maximum removal; physical audio devices excluded.",
    "C++ and NumPy use different RNG implementations; identical numeric seeds do not mean identical latent noise.",
    "Memory includes the Python/NumPy measurement host; runtime inference has no Python requirement.",
    "Loona reference is the same 1.05s WAV used by the previous current-Sherpa test. Official PyTorch baseline used Alba preset state, so it is not voice-identical.",
    "Files named cpp-mixed-alba-raw and cpp-mixed-alba-gate50 in the initial matrix were mislabelled: their voice is the Expresso ex03-ex01_happy_001_channel1_334s.wav reference. Report labels correct this. The later cpp-mixed-alba3-gate50 uses actual alba-mackenna/casual.wav, first 3 seconds, resampled to 24 kHz.",
    "Additional prepared-voice runs opt into disk embedding/KV reuse under separate per-study directories. One preparation process is explicitly measured separately.",
    "Final quality transcription is not CPU-pinned; only TTS inference is timed and pinned. Recognizer model/settings remain the same.",
    "Comparing separately generated raw and filtered WAVs does not prove exact PCM parity: pipelining can alter final decoder chunk sizes and floating-point/quantization output. Unit tests prove the onset filter itself returns an exact suffix across variable chunk boundaries."
  ]
}