# CosyVoice ONNX Artifacts — PROVENANCE

## Source

- **Upstream repo**: https://github.com/FunAudioLLM/CosyVoice
- **Model weights**: https://huggingface.co/FunAudioLLM/CosyVoice-300M
- **License**: Apache-2.0 (code and weights)

## License note

The Emilia training corpus is CC-BY-NC-4.0.  The **model weights** are
independently licensed under Apache-2.0 by the authors; the data license does
not restrict downstream model use.  This is the same legal position as most
foundation model releases.

## Non-AR VC recipe

```
speech_tokenizer_v1.onnx(source_whisper_mel) → content_tokens
campplus.onnx(reference_fbank)               → speaker_embedding (192d)
flow_encoder.onnx(content_tokens)            → mu (1, 80, T_mel)
flow_decoder.onnx(mu, spk_embed, 10 ODE steps) → mel (1, 80, T_mel)
hifigan_f0_source.onnx(mel)                  → source_1d (NSF harmonic)
numpy_stft(source_1d, n_fft=16, hop=4)       → source_stft (1, 18, T_stft)
hifigan_backbone.onnx(mel, source_stft)      → (magnitude, phase)
numpy_istft(magnitude, phase, n_fft=16)      → waveform @ 22050 Hz
```

The autoregressive LLM (llm.pt) is not used.

## Components

| File | Origin | Description |
|---|---|---|
| `speech_tokenizer_v1.onnx` | upstream HF | SenseVoice content tokenizer |
| `campplus.onnx` | upstream HF | CAM++ speaker encoder (192-d) |
| `flow_decoder.onnx` | upstream HF | Flow-matching ODE decoder |
| `flow_encoder.onnx` | exported | Token embed + 6-block Conformer → mu |
| `hifigan_f0_source.onnx` | exported | F0 predictor + NSF source generator |
| `hifigan_backbone.onnx` | exported | HiFiGAN conv decoder → STFT coefficients |

## STFT/ISTFT note

`aten::stft` / `aten::istft` are not supported at ONNX opset 14.  The HiFiGAN
is split at the STFT boundary; the adapter implements numpy STFT/ISTFT
(n_fft=16, hop_len=4, Hann window) verified at export time against PyTorch.

## Export toolchain

`conversion/export_cosyvoice.py` in TigreGotico/voiceclonnx (Apache-2.0)