config.json

{
  "engine": "omnivoice",
  "model_type": "omnivoice",
  "upstream": "k2-fsa/OmniVoice",
  "architecture": "masked-diffusion (discrete-diffusion) codec LM; Qwen3-0.6B backbone, bidirectional attention",
  "audio": {"sample_rate": 24000},
  "frame_rate": 25,
  "hop_length": 960,
  "num_audio_codebook": 8,
  "audio_vocab_size": 1025,
  "audio_mask_id": 1024,
  "audio_codebook_weights": [8, 8, 6, 6, 4, 4, 2, 2],
  "codec": "Higgs Audio V2 tokenizer (DAC acoustic + HuBERT semantic, 8 RVQ codebooks)",
  "graphs": {
    "model": "omnivoice_backbone.onnx",
    "acoustic_encoder_path": "acoustic_encoder.onnx",
    "semantic_encoder_path": "semantic_encoder.onnx",
    "quantizer_encoder_path": "quantizer_encoder.onnx",
    "decoder_path": "higgs_decoder.onnx",
    "bpe_tokenizer_path": "tokenizer.json"
  },
  "generation_defaults": {
    "num_step": 32,
    "guidance_scale": 2.0,
    "t_shift": 0.1,
    "layer_penalty_factor": 5.0,
    "position_temperature": 5.0,
    "class_temperature": 0.0
  }
}