README.md

December 18, 2025 ยท View on GitHub

\Large \boldsymbol{\mathsf{\color{#6366f1}M\color{#a855f7}3\color{#ec4899}\text{-}TTS}}: \mathsf{\color{#de2910}M\color{black}\text{ulti-}\color{#de2910}M\color{black}\text{odal\ DiT\ Alignment}\ \color{black}\\&\ \color{#de2910}M\color{black}\text{el-latent}}

arXiv Demo Page

๐Ÿ“… Roadmap

  • Release model code
  • Release training and inference code
  • Release pre-trained model weights

๐Ÿ”ฅ Key Features

  • No Pseudo-Alignment: Achieves stable alignment implicitly via Joint-DiT attention.
  • Mel-VAE Codec: Efficient latent representation for faster training and high-fidelity reconstruction.
  • Unified Architecture: A simple, end-to-end framework without complex multi-stage pipelines.

๐Ÿ™Œ Acknowledgements

This project is built upon the excellent work of F5-TTS, MMAudio and Zip-Voice. We thank the authors for their open-source contributions.

๐Ÿ“ Citation

If you find our work helpful for your research, please consider giving a star โญ and citation ๐Ÿ“:

@article{wang2025m3tts,
  title={M3-TTS: Multi-modal DiT Alignment \& Mel-latent for Zero-shot High-fidelity Speech Synthesis},
  author={Wang, Xiaopeng and Qiang, Chunyu and Fu, Ruibo and others},
  journal={arXiv preprint arXiv:2512.04720},
  year={2025}
}