🚀 TaDiCodec

We introduce the Text-aware Diffusion Transformer Speech Codec (TaDiCodec), a novel approach to speech tokenization that employs end-to-end optimization for quantization and reconstruction through a diffusion autoencoder, while integrating text guidance into the diffusion decoder to enhance reconstruction quality and achieve optimal compression. TaDiCodec achieves an extremely low frame rate of 6.25 Hz and a corresponding bitrate of 0.0875 kbps with a single-layer codebook for 24 kHz speech, while maintaining superior performance on critical speech generation evaluation metrics such as Word Error Rate (WER), speaker similarity (SIM), and speech quality (UTMOS).

🤗 Pre-trained Models

🔧 Quick Model Usage

# 🤗 Load from Hugging Face
from models.tts.tadicodec.inference_tadicodec import TaDiCodecPipline
from models.tts.llm_tts.inference_llm_tts import TTSInferencePipeline
from models.tts.llm_tts.inference_mgm_tts import MGMInferencePipeline

# Load TaDiCodec tokenizer, it will automatically download the model from Hugging Face for the first time
tokenizer = TaDiCodecPipline.from_pretrained("amphion/TaDiCodec")

# Load AR TTS model, it will automatically download the model from Hugging Face for the first time  
tts_model = TTSInferencePipeline.from_pretrained("amphion/TaDiCodec-TTS-AR-Qwen2.5-3B")

# Load MGM TTS model, it will automatically download the model from Hugging Face for the first time
tts_model = MGMInferencePipeline.from_pretrained("amphion/TaDiCodec-TTS-MGM")

🚀 Quick Start

Installation

# Clone the repository
git clone https://github.com/HeCheng0625/Diffusion-Speech-Tokenizer.git
cd Diffusion-Speech-Tokenizer

# Install dependencies
bash env.sh

Basic Usage

Please refer to the use_examples folder for more detailed usage examples.

Speech Tokenization and Reconstruction

# Example: Using TaDiCodec for speech tokenization
import torch
import soundfile as sf
from models.tts.tadicodec.inference_tadicodec import TaDiCodecPipline

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
pipe = TaDiCodecPipline.from_pretrained(ckpt_dir="./ckpt/TaDiCodec", device=device)

# Text of the prompt audio
prompt_text = "In short, we embarked on a mission to make America great again, for all Americans."
# Text of the target audio
target_text = "But to those who knew her well, it was a symbol of her unwavering determination and spirit."

# Input audio path of the prompt audio
prompt_speech_path = "./use_examples/test_audio/trump_0.wav"
# Input audio path of the target audio
speech_path = "./use_examples/test_audio/trump_1.wav"

rec_audio = pipe(
    text=target_text,
    speech_path=speech_path,
    prompt_text=prompt_text,
    prompt_speech_path=prompt_speech_path
)
sf.write("./use_examples/test_audio/trump_rec.wav", rec_audio, 24000)

Zero-shot TTS with TaDiCodec

import torch
import soundfile as sf
from models.tts.llm_tts.inference_llm_tts import TTSInferencePipeline
# from models.tts.llm_tts.inference_mgm_tts import MGMInferencePipeline

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# Create AR TTS pipeline
pipeline = TTSInferencePipeline.from_pretrained(
    tadicodec_path="./ckpt/TaDiCodec",
    llm_path="./ckpt/TaDiCodec-TTS-AR-Qwen2.5-3B",
    device=device,
)

# Inference on single sample, you can also use the MGM TTS pipeline
audio = pipeline(
    text="但是 to those who 知道 her well, it was a 标志 of her unwavering 决心 and spirit.",   # code-switching cases are supported
    prompt_text="In short, we embarked on a mission to make America great again, for all Americans.",
    prompt_speech_path="./use_examples/test_audio/trump_0.wav",
)

sf.write("./use_examples/test_audio/lm_tts_output.wav", audio, 24000)

📚 Citation

If you find this repository useful, please cite our paper:

TaDiCodec:

@article
	{tadicodec2025,
  title={TaDiCodec: Text-aware Diffusion Speech Tokenizer for Speech Language Modeling},
  author={Yuancheng Wang, Dekun Chen, Xueyao Zhang, Junan Zhang, Jiaqi Li, Zhizheng Wu},
  journal={arXiv preprint},
  year={2025},
  url={https://arxiv.org/abs/2508.16790}
}

Amphion:

@inproceedings{amphion,
    author={Xueyao Zhang and Liumeng Xue and Yicheng Gu and Yuancheng Wang and Jiaqi Li and Haorui He and Chaoren Wang and Ting Song and Xi Chen and Zihao Fang and Haopeng Chen and Junan Zhang and Tze Ying Tang and Lexiao Zou and Mingxuan Wang and Jun Han and Kai Chen and Haizhou Li and Zhizheng Wu},
    title={Amphion: An Open-Source Audio, Music and Speech Generation Toolkit},
    booktitle={{IEEE} Spoken Language Technology Workshop, {SLT} 2024},
    year={2024}
}

MaskGCT:

@inproceedings{wang2024maskgct,
  author={Wang, Yuancheng and Zhan, Haoyue and Liu, Liwei and Zeng, Ruihong and Guo, Haotian and Zheng, Jiachen and Zhang, Qiang and Zhang, Xueyao and Zhang, Shunsi and Wu, Zhizheng},
  title={MaskGCT: Zero-Shot Text-to-Speech with Masked Generative Codec Transformer},
  booktitle    = {{ICLR}},
  publisher    = {OpenReview.net},
  year         = {2025}
}

🙏 Acknowledgments

• MGM-based TTS is built upon MaskGCT.

• Vocos vocoder is built upon Vocos.

• NAR Llama-style transformers is built upon transformers.

• (Binary Spherical Quantization) BSQ is built upon vector-quantize-pytorch and bsq-vit.

• Training codebase is built upon Amphion and accelerate.