Model Card for UniSS

Model Details

Model Description

UniSS is a unified single-stage speech-to-speech translation (S2ST) framework that achieves high translation fidelity and speech quality, while preserving timbre, emotion, and duration consistency. UniSS supports English and Chinese now.

Model Sources

• Repository: https://github.com/cmots/UniSS
• Paper: https://arxiv.org/pdf/2509.21144
• Demo: https://cmots.github.io/uniss-demo

Quick Start

1. Install the environment and get the code

conda create -n uniss python=3.10.16
conda activate uniss
git clone https://github.com/cmots/UniSS.git
cd UniSS
pip install -r requirements.txt
# If you are in mainland China, you can set the mirror as follows:
pip install -r requirements.txt -i https://mirrors.aliyun.com/pypi/simple/ --trusted-host=mirrors.aliyun.com

2. Download the weight

The weight of UniSS is on HuggingFace.

You have to download the model manually, you can download it via provided script:

python download_weight.py

or download via git clone (skip this if you have download via python script):

mkdir -p pretrained_models

# Make sure you have git-lfs installed (https://git-lfs.com)
git lfs install

git clone https://huggingface.co/cmots/UniSS pretrained_models/UniSS

3. Run the code

import soundfile
from uniss import UniSSTokenizer
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
from uniss import process_input, process_output

# 1. Set the device, wav path, model path
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

wav_path = "prompt_audio.wav"
model_path = "pretrained_models/UniSS"

# 2. Set the mode and target language
mode = 'Quality'    # 'Quality' or 'Performance'
tgt_lang = "<|eng|>"    # for English output
# tgt_lang = "<|cmn|>"  # for Chinese output

# 3. load the model, text tokenizer, and speech tokenizer
model = AutoModelForCausalLM.from_pretrained(model_path, device_map=device)
tokenizer = AutoTokenizer.from_pretrained(model_path)

speech_tokenizer = UniSSTokenizer.from_pretrained(model_path, device=device)

# 4. extract speech tokens
glm4_tokens, bicodec_tokens = speech_tokenizer.tokenize(wav_path)


# 5. process the input
input_text = process_input(glm4_tokens, bicodec_tokens, mode, tgt_lang)
input_token_ids = tokenizer.encode(input_text, return_tensors="pt").to(device)

# 6. translate the speech
output = model.generate(
    input_token_ids,
    max_new_tokens=1500,
    temperature=0.7,
    top_p=0.8,
    repetition_penalty=1.1
)

# 7. decode the output
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)

# 8. process the output
audio, translation, transcription = process_output(output_text[0], input_text, speech_tokenizer, mode, device)

# 9. save and show the results
soundfile.write("output_audio.wav", audio, 16000)

if mode == 'Quality':
    print("Transcription:\n", transcription)
print("Translation:\n", translation)

More examples and details is on Our Github Repo.

Citation

If you find our paper and code useful in your research, please consider giving a like and citation.

@misc{cheng2025uniss_s2st,
      title={UniSS: Unified Expressive Speech-to-Speech Translation with Your Voice}, 
      author={Sitong Cheng and Weizhen Bian and Xinsheng Wang and Ruibin Yuan and Jianyi Chen and Shunshun Yin and Yike Guo and Wei Xue},
      year={2025},
      eprint={2509.21144},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2509.21144}, 
}