---
library_name: pytorch
license: mit
pipeline_tag: text-to-speech
tags:
- accent-tts
- mandarin
- joycent
- grad-tts
- whisaid
---

# Joycent trained with WhisAID Medium GRL accent embeddings

This is a Joycent Mandarin accent TTS acoustic model trained using accent
embeddings extracted by
[`walston/whisaid-medium-grl`](https://huggingface.co/walston/whisaid-medium-grl).
The released checkpoint is epoch 100.

- **Paper:** [Joycent: Diffusion-based Accent TTS without Accented Phone Prediction](https://huggingface.co/papers/2606.16417)
- **Code:** [oshindow/Joycent-code](https://github.com/oshindow/Joycent-code)
- **Accent encoder:** [walston/whisaid-medium-grl](https://huggingface.co/walston/whisaid-medium-grl)
- **Vocoder:** [walston/joycent-vocoder](https://huggingface.co/walston/joycent-vocoder)

## Download

```python
from huggingface_hub import hf_hub_download

checkpoint_path = hf_hub_download(
    repo_id="walston/joycent-medium-grl",
    filename="grad_100.pt",
)
```

Pass the downloaded checkpoint to `joycent/inference_joycent.py` with the
`--acoustic-checkpoint` argument. Full synthesis also requires the Joycent
vocoder and reference-audio feature extraction dependencies described in the
Joycent repository.

## Checkpoint

- Epoch: 100
- Acoustic model: Joycent / Grad-TTS
- Accent embedding model: WhisAID Whisper Medium GRL (lambda 0.05)
- Accent embedding dimension: 256

## Citation

```bibtex
@misc{wang2026joycentdiffusionbasedaccenttts,
      title={Joycent: Diffusion-based Accent TTS without Accented Phone Prediction},
      author={Xintong Wang and Ye Wang},
      year={2026},
      eprint={2606.16417},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
}
```