---
license: cc-by-nc-4.0
pipeline_tag: text-to-speech
tags:
- text-to-speech
- tts
- voice-cloning
- flow-matching
- raw-waveform
- english
- chinese
- wavtts
- mirror
---

# WavTTS — inference mirror for MAESTRO

Inference-ready repack of
[worstchan/WavTTS](https://huggingface.co/worstchan/WavTTS)
(Wenxi Chen et al., [arXiv 2606.03455](https://arxiv.org/abs/2606.03455)),
kept by AEmotionStudio so the MAESTRO DAW can fetch the checkpoint on demand.
The model is **not** fine-tuned, quantised below its training precision, or
otherwise altered — this repo only extracts the inference weights from the
upstream training checkpoint:

* `model_1200000.pt` (10.8 GB training checkpoint) → `model.safetensors`:
  the `ema_model_state_dict` EMA weights with the `ema_model.` prefix
  stripped, EMA bookkeeping (`initted`/`step`) and the training-only
  `aux_mel_loss.*` buffers dropped, floating tensors stored as float16
  (the dtype upstream inference uses on CUDA).
* `vocab.txt` — verbatim from upstream (2556-entry pinyin/char vocab).
* `config.json` — pointer config updated to reference `model.safetensors`.

| File | Purpose |
|---|---|
| `model.safetensors` | ~0.67B DiT + CFM inference weights (fp16 EMA) |
| `vocab.txt` | Character/pinyin vocabulary (index 0 = space/unknown) |
| `config.json` | Sample rate + file pointers |

WavTTS generates the **raw 16 kHz waveform directly** (flow matching over
160-sample waveform patches — no mel spectrogram, VAE latent or codec).
Zero-shot voice cloning from a reference clip plus its transcript; English
and Chinese.

## License — please read

* **Code**: MIT (the [WavTTS codebase](https://github.com/cwx-worst-one/WavTTS)).
* **Weights (this repo): CC BY-NC 4.0** — inherited from the Emilia training
  dataset. **Non-commercial use only**; attribution required. Do not use
  outputs commercially without permission from the upstream authors.
* Never clone or impersonate a real person's voice without their explicit
  consent.

By downloading these files you accept the CC BY-NC 4.0 terms. MAESTRO shows
the same terms in-app.

## Citation

```bibtex
@article{chen2026wavtts,
  title   = {WavTTS: Towards High-Quality Zero-Shot TTS via Direct Raw Waveform Modeling},
  author  = {Chen, Wenxi and Jia, Dongya and Chen, Yushen and Niu, Zhikang and Liang, Yuzhe and Li, Xiquan and Yan, Ruiqi and Ma, Ziyang and Yang, Guanrou and Chen, Sanyuan and others},
  journal = {arXiv preprint arXiv:2606.03455},
  year    = {2026},
}
```