{
  "schema": "transcribe-golden-manifest-v1",
  "family": "voxtral_realtime",
  "variant": "voxtral-mini-4b-realtime-2602",
  "source_model": {
    "hf_repo": "mistralai/Voxtral-Mini-4B-Realtime-2602",
    "hf_revision": "2769294da9567371363522aac9bbcfdd19447add"
  },
  "reference": {
    "kind": "transformers",
    "source": "https://github.com/huggingface/transformers",
    "revision": "v5.10.2",
    "entrypoint": "scripts/dump_reference_voxtral_realtime_transformers.py",
    "dump_args": [],
    "notes": "Offline whole-clip oracle is the per-tensor contract (decode subcommand). A separate streaming-replay pass (stream subcommand) dumps the streaming transcript + per-step logits so Stage 4 can verify offline<->streaming numerical equivalence (intake risk #1). num_delay_tokens defaults to the processor value (config default 6 = 480ms, the publisher-recommended setting). The processor is auto-detect only (TranscriptionRequest language=None); --language is accepted but ignored."
  },
  "expected_dtype": "bfloat16",
  "dtype_source": "config",
  "frontend": {
    "sample_rate": 16000,
    "n_mels": 128,
    "hop_length": 160,
    "fft_size": 400,
    "win_length": 400,
    "window": "hann_periodic",
    "normalization": "global",
    "preemphasis": null,
    "dither": 0.0
  },
  "tokenizer_summary": {
    "type": "bpe",
    "vocab_size": 131072,
    "special_tokens": {
      "bos": 1,
      "eos": 2,
      "pad": 11,
      "streaming_pad": 32
    }
  },
  "capabilities": {
    "languages": ["en", "fr", "es", "de", "ru", "zh", "ja", "it", "pt", "nl", "ar", "hi", "ko"],
    "language_detection": true,
    "translation": false,
    "timestamps": ["none"],
    "streaming": true,
    "speaker_diarization": false
  },
  "tolerance_file": "tests/tolerances/voxtral_realtime.json",
  "cases": ["jfk"]
}
