Gwen-TTS 0.6B - Natural Vietnamese Voice Cloning

Gwen-TTS is a Vietnamese text-to-speech model with natural voice cloning capability.

Key highlights:

• Clone any voice with just a few seconds of reference audio
• Natural and expressive Vietnamese voice cloning
• Finetuned from Qwen3-TTS-0.6B on ~1,000 hours of Vietnamese audio data crawled from TikTok

Demo: https://g-voice.g-ailab.com/tts (integrated with TTS text normalization and serving)

GitHub: https://github.com/ggroup-ai-lab/gwen-tts

How to Use

Installation

pip install -U qwen-tts
# Optional: for optimized performance
pip install -U flash-attn --no-build-isolation

Voice Cloning

Note: For best quality, proactively apply TTS text normalization (numbers, symbols, abbreviations, etc.) and split input into chunks before passing to the model.

import torch
import soundfile as sf
from qwen_tts import Qwen3TTSModel

# Load model
model = Qwen3TTSModel.from_pretrained(
    "g-group-ai-lab/gwen-tts-0.6B",
    device_map="cuda:0",
    dtype=torch.bfloat16,
    attn_implementation="flash_attention_2",
)

# Recommended generation config for Gwen-TTS
generation_config = dict(
    temperature=0.3,
    top_k=20,
    top_p=0.9,
    max_new_tokens=4096,
    repetition_penalty=2.0,
    subtalker_do_sample=True,
    subtalker_temperature=0.1,
    subtalker_top_k=20,
    subtalker_top_p=1.0,
)

# Generate speech with voice cloning
wavs, sr = model.generate_voice_clone(
    text="<your text to synthesize>",
    language="Vietnamese",
    ref_audio="<path/to/reference.wav>",
    ref_text="<transcript of the reference audio>",
    **generation_config,
)

sf.write("output.wav", wavs[0], sr)

Voice Samples

Each speaker shows the reference audio (voice input) and the generated inference audio (output).

Supported Languages

Vietnamese (primary), Chinese, English, Japanese, Korean, French, German, Italian, Portuguese, Russian, Spanish.

Note: This model is optimized for Vietnamese. Performance on other languages may differ from the base Qwen3-TTS model.

Citation

@misc{gwen-tts,
    title={Gwen-TTS: Natural Vietnamese Voice Cloning},
    author={G-Group AI Lab},
    year={2026},
    url={https://github.com/ggroup-ai-lab/gwen-tts}
}

License

This model is released under the MIT License.

Acknowledgments

• Qwen Team for the Qwen3-TTS base model
• G-Group AI Lab for training and releasing this model