{
  "engine": "cosyvoice",
  "components": {
    "speech_tokenizer": "speech_tokenizer_v1.onnx",
    "campplus": "campplus.onnx",
    "flow_encoder": "flow_encoder.onnx",
    "flow_encoder_q8": "flow_encoder_q8.onnx",
    "flow_decoder": "flow_decoder.onnx",
    "flow_decoder_q8": "flow_decoder_q8.onnx",
    "hifigan_f0_source": "hifigan_f0_source.onnx",
    "hifigan_f0_source_q8": "hifigan_f0_source_q8.onnx",
    "hifigan_backbone": "hifigan_backbone.onnx",
    "hifigan_backbone_q8": "hifigan_backbone_q8.onnx"
  },
  "sample_rates": {
    "input": 16000,
    "output": 22050
  },
  "opset": 14,
  "upstream_repo": "FunAudioLLM/CosyVoice-300M",
  "license": "Apache-2.0",
  "vc_recipe": "Non-AR VC: speech_tokenizer(src@16kHz) \u2192 content_tokens; campplus(ref@16kHz) \u2192 spk_emb(192); flow_encoder(tokens) \u2192 mu(1,80,T_mel); flow_decoder(mu, spk_emb, ODE 10 steps) \u2192 mel(1,80,T_mel); hifigan_f0_source(mel) \u2192 src_1d; numpy_stft(src_1d) \u2192 src_stft; hifigan_backbone(mel, src_stft) \u2192 (mag, phase); numpy_istft(mag, phase) \u2192 waveform@22050Hz. No LLM.",
  "cpu_rtf_estimate": 0.71,
  "hifigan_stft_params": {
    "n_fft": 16,
    "hop_len": 4
  }
}