{
  "model": "speaker_feature_extractor",
  "runtime": "TFLite builtins",
  "source": "SeamlessM4TFeatureExtractor deterministic fbank preprocessing",
  "sample_rate": 16000,
  "input_contract": {
    "name": "waveform",
    "shape": [
      1,
      64000
    ],
    "dtype": "float32",
    "note": "4-second 16 kHz normalized mono waveform; shorter clips are padded by the caller."
  },
  "output_contract": {
    "name": "input_features",
    "shape": [
      1,
      199,
      160
    ],
    "dtype": "float32",
    "note": "Matches Wav2Vec2-BERT/SeamlessM4T feature stacking expected by speaker_features_to_embedding.tflite."
  },
  "preprocessing": {
    "frame_length": 400,
    "hop_length": 160,
    "fft_length": 512,
    "num_mel_bins": 80,
    "stride": 2,
    "preemphasis": 0.97,
    "window": "povey(periodic=False)",
    "mel_scale": "kaldi",
    "mel_floor": 1.1920928955078125e-07,
    "normalization": "per-mel zero mean, sample variance ddof=1"
  },
  "output": {
    "path": "speaker_feature_extractor.tflite",
    "bytes": 98848,
    "sha256": "9eb3d8c19df7b508fd66bd18206c4c4b36fd0fe85ce91c3075057c4d1400a5db"
  },
  "verification": {
    "enabled": true,
    "reference": "SeamlessM4TFeatureExtractor reference output",
    "input_shape": [
      1,
      64000
    ],
    "output_shape": [
      1,
      199,
      160
    ],
    "max_abs_diff": 5.7220458984375e-05,
    "mean_abs_diff": 5.468903054861585e-06,
    "p99_abs_diff": 2.6464462280273438e-05
  }
}