config.json

{
  "name": "V1: EmbeddingGemma 2 audio tower (layers 0-11) + causal CTC head, Core ML",
  "version": "v1",
  "license": "apache-2.0",
  "base_model": {
    "id": "google/embeddinggemma-2",
    "revision": "914f7f89142e33e77833254d9c9b90c3cef7303b",
    "license": "apache-2.0",
    "component": "audio_tower",
    "layers": [
      0,
      11
    ]
  },
  "package": {
    "file": "v1.mlpackage",
    "format": "mlprogram (multifunction, shared weights)",
    "precision": "fp16",
    "source_build": "fp16safe/fp16",
    "specification_version": 9,
    "minimum_os": "macOS 15 / iOS 18",
    "compute_units": {
      "recommended": "CPU_AND_NE",
      "supported": [
        "CPU_AND_NE",
        "CPU_AND_GPU",
        "CPU_ONLY"
      ]
    },
    "default_function": "win_10s"
  },
  "windows": [
    {
      "function": "win_5s",
      "seconds": 5.0,
      "mel_frames": 500,
      "output_frames": 125
    },
    {
      "function": "win_10s",
      "seconds": 10.0,
      "mel_frames": 1000,
      "output_frames": 250
    },
    {
      "function": "win_20s",
      "seconds": 20.0,
      "mel_frames": 2000,
      "output_frames": 500
    },
    {
      "function": "win_35s",
      "seconds": 35.0,
      "mel_frames": 3500,
      "output_frames": 875
    }
  ],
  "window_choice": "smallest window with mel_frames >= valid mel frames",
  "inputs": [
    {
      "name": "mel",
      "shape": [
        "1",
        "T_mel",
        "128"
      ],
      "dtype": "float32",
      "description": "log-mel frames (see feature_extractor), zero-padded to the function's window"
    },
    {
      "name": "valid_len",
      "shape": [
        "1"
      ],
      "dtype": "int32",
      "description": "number of valid mel frames"
    }
  ],
  "outputs": [
    {
      "name": "log_probs",
      "shape": [
        "1",
        "T_mel/4",
        "257"
      ],
      "dtype": "float16",
      "description": "CTC log-probabilities (natural log), 256 pieces + blank"
    },
    {
      "name": "l11",
      "shape": [
        "1",
        "T_mel/4",
        "1024"
      ],
      "dtype": "float16",
      "description": "layer-11 features after the int8 round trip"
    }
  ],
  "valid_output_frames": "ceil(valid_len / 4)",
  "feature_extractor": {
    "class": "Gemma4AudioFeatureExtractor",
    "preprocessor_config": {
      "dither": 0.0,
      "feature_extractor_type": "Gemma4AudioFeatureExtractor",
      "feature_size": 128,
      "fft_length": 512,
      "fft_overdrive": false,
      "frame_length": 320,
      "hop_length": 160,
      "input_scale_factor": 1.0,
      "max_frequency": 8000.0,
      "mel_floor": 0.001,
      "min_frequency": 0.0,
      "padding_side": "right",
      "padding_value": 0.0,
      "per_bin_mean": null,
      "per_bin_stddev": null,
      "preemphasis": 0.0,
      "preemphasis_htk_flavor": true,
      "return_attention_mask": true,
      "sampling_rate": 16000
    },
    "source": "google/embeddinggemma-2@914f7f89142e33e77833254d9c9b90c3cef7303b/preprocessor_config.json (unchanged)",
    "call": {
      "sampling_rate": 16000,
      "padding": "longest",
      "truncation": false
    },
    "valid_mel_frames": "sum(input_features_mask)",
    "mel_frame_rate_hz": 100
  },
  "frame_rate_hz": 25,
  "subsampling": 4,
  "lookahead_ms": 160,
  "encoder_lookahead_ms": 0,
  "ctc_head": {
    "params": 13283585,
    "future_frames": 4,
    "causal_conv": true
  },
  "vocab": {
    "tokenizer": "spm_unigram256.model",
    "type": "sentencepiece-unigram",
    "size": 256,
    "blank_id": 256,
    "num_classes": 257
  },
  "decoder": {
    "library": "ctc-beam-decoder",
    "library_version": "0.2.0",
    "lm": "sp256_6gram.bin",
    "lm_order": 6,
    "piece_word_prefix": "t",
    "alpha": 0.8,
    "beta": 1.0,
    "beam": 32,
    "units": "piece",
    "beam_prune_logp": -10.0,
    "token_min_logp": -5.0,
    "unk_score_offset": -10.0,
    "lm_score_boundary": true,
    "prune_history": true
  },
  "long_audio": {
    "max_mel_frames": 3500,
    "left_context_s": 2.0,
    "cut": "quietest point in the last 10 s (code repo: coreml/transcribe.py)"
  },
  "text_normalization": "lowercase, collapse whitespace"
}