DELTA-TTS

     

DELTA-TTS converts a pretrained autoregressive TTS model into a discrete diffusion language model with LoRA and a convolution module, so that speech tokens are decoded in confidence order instead of left to right.

• Paper: DELTA-TTS: Adapting Autoregressive Model into Diffusion Language Model for Text-to-Speech
• Repository: GitHub
• Demo: Demo Page

Files

File	Contents
delta_tts.pt	DELTA-TTS weights: LoRA (r=64), depth-wise convolution modules (k=31) and the [MASK] embedding, 94M parameters

The backbone, speech tokenizer, flow-matching decoder and vocoder are those of Fun-CosyVoice3-0.5B-2512 and are loaded from that repository.

Usage

Installation and inference are described in the GitHub repository.

huggingface-cli download FunAudioLLM/Fun-CosyVoice3-0.5B-2512 --local-dir pretrained_models/Fun-CosyVoice3-0.5B-2512
huggingface-cli download slpr-lab/DELTA-TTS --local-dir pretrained_models/DELTA-TTS

Citation

@inproceedings{moon2026deltatts,
  title     = {{DELTA-TTS}: Adapting Autoregressive Model into Diffusion Language Model for Text-to-Speech},
  author    = {Moon, Junwon and Lee, Yejin and Kim, Seungbeom and Ahn, Hoseong and Park, Sewoong and Kim, Heeseung and Shim, Kyuhong},
  booktitle = {Advances in Neural Information Processing Systems},
  year      = {2026}
}

License

The weights are released under the Apache-2.0 license. They are built on Fun-CosyVoice3-0.5B-2512, which is also released under the Apache-2.0 license.