{
  "model_id": "Prince-1/OmniVoice",
  "execution_provider": "CPUExecutionProvider",
  "precision": "int4",
  "backbone": {
    "audio_embeddings_encoder": {
      "filename": "audio_embeddings_encoder.onnx",
      "description": "Fuses text + audio codec token embeddings \u2192 inputs_embeds",
      "inputs": {
        "input_ids": "int64 (batch, num_codebooks=8, seq)",
        "audio_mask": "bool  (batch, seq)"
      },
      "outputs": {
        "inputs_embeds": "float32 (batch, seq, hidden=1024)"
      }
    },
    "llm_decoder": {
      "filename": "llm_decoder.onnx",
      "description": "Qwen3 28-layer backbone (inputs_embeds \u2192 hidden_states)",
      "note": "Exported with exclude_embeds=True, exclude_lm_head=True",
      "inputs": {
        "inputs_embeds": "float32 (batch, seq, hidden=1024)",
        "attention_mask": "int64   (batch, seq)",
        "position_ids": "int64   (batch, seq)",
        "past_key_values": "float32 per layer \u2014 pass empty (shape [B,heads,0,head_dim]) for full-sequence forward"
      },
      "outputs": {
        "hidden_states": "float32 (batch, seq, hidden=1024)"
      },
      "hidden_size": 1024,
      "num_layers": 28,
      "num_attn_heads": 16,
      "num_kv_heads": 8,
      "head_dim": 128
    },
    "audio_heads_decoder": {
      "filename": "audio_heads_decoder.onnx",
      "description": "Projects hidden_states to per-codebook audio-token logits",
      "inputs": {
        "hidden_states": "float32 (batch, seq, hidden=1024)"
      },
      "outputs": {
        "logits": "float32 (batch, num_codebooks=8, seq, audio_vocab=1025)"
      },
      "num_codebooks": 8,
      "audio_vocab_size": 1025,
      "audio_mask_id": 1024
    }
  },
  "higgs_tokenizer": {
    "note": "Exported via optimize.py --include-higgs (or --higgs-only).",
    "models_dir": "higgs/models/",
    "sub_models": {
      "acoustic_encoder": "acoustic_encoder.onnx",
      "semantic_encoder": "semantic_encoder.onnx",
      "quantizer_encoder": "quantizer_encoder.onnx",
      "higgs_decoder": "higgs_decoder.onnx"
    },
    "sample_rate_acoustic": 24000,
    "sample_rate_semantic": 16000,
    "downsample_factor": 320,
    "num_codebooks": 8,
    "codebook_size": 1024,
    "pipeline": [
      "acoustic_encoder(waveform_24k) \u2192 acoustic_features",
      "semantic_encoder(waveform_16k) \u2192 semantic_features",
      "quantizer_encoder(acoustic_features, semantic_features) \u2192 codes",
      "  --- TTS inference: codes \u2192 audio_embeddings_encoder \u2192 ... \u2192 audio_codes ---",
      "higgs_decoder(audio_codes) \u2192 waveform_24k"
    ],
    "loading": "transformers.AutoModel (transformers>=5.4.0, no external deps needed)",
    "docs": "https://huggingface.co/docs/transformers/v5.4.0/en/model_doc/higgs_audio_v2_tokenizer"
  },
  "iterative_decoding": {
    "note": "OmniVoice uses 32-step iterative unmasking (non-autoregressive).",
    "steps": 32,
    "audio_codebook_weights": [
      8,
      8,
      6,
      6,
      4,
      4,
      2,
      2
    ],
    "per_step_pipeline": [
      "audio_embeddings_encoder(input_ids, audio_mask) \u2192 inputs_embeds",
      "llm_decoder(inputs_embeds, attention_mask, position_ids) \u2192 hidden_states",
      "audio_heads_decoder(hidden_states) \u2192 logits",
      "sample audio tokens from logits using codebook weights",
      "unmask predicted positions in input_ids"
    ]
  }
}