Qwen3-TTS-TBLive-Base
A Chinese live-streaming domain TTS checkpoint fine-tuned from
Qwen3-TTS-12Hz-1.7B-Base via continual pre-training (CPT) and GRPO
post-training on live-streaming speech data.
Recommended starting point for downstream fine-tuning. If you need production-ready built-in live-streaming voices, use the sibling checkpoint TaoLiveAIGC/Qwen3-TTS-TBLive-CustomVoice.
Benchmark
Zero-shot voice-cloning results on the public
seed-tts-eval test sets,
compared against open-source baselines. / marks numbers not reported.
| Model | test-en SIM-o ↑ | test-en WER ↓ | test-en UTMOS ↑ | test-zh SIM-o ↑ | test-zh WER ↓ | test-zh UTMOS ↑ |
|---|---|---|---|---|---|---|
| Ground-truth | 0.734 | 2.14 | 3.52 | 0.755 | 1.25 | 2.78 |
| IndexTTS2 | 0.706 | 2.33 | 3.65 | 0.764 | 1.05 | 3.00 |
| CosyVoice3 | 0.696 | 2.17 | 3.96 | 0.778 | 1.14 | 3.32 |
| VoxCPM | 0.731 | 1.92 | 3.77 | 0.772 | 0.99 | 2.94 |
| MossTTS Local | 0.732 | 1.93 | / | 0.796 | 1.44 | / |
| Qwen3-TTS | 0.708 | 1.54 | 4.16 | 0.766 | 1.15 | 3.46 |
| Qwen3-TTS-TBLive-Base | 0.732 | 1.63 | 4.11 | 0.782 | 1.28 | 3.38 |
Quickstart
pip install -U qwen-tts
import torch
from qwen_tts import Qwen3TTSModel
model = Qwen3TTSModel.from_pretrained(
"TaoLiveAIGC/Qwen3-TTS-TBLive-Base",
device_map="cuda:0",
dtype=torch.bfloat16,
)
# Zero-shot voice cloning with a short reference audio
wavs, sr = model.generate_voice_clone(
text="家人们，今天直播间的福利真的拉满了！",
language="Chinese",
ref_audio="path/to/reference.wav",
ref_text="参考音频对应的文本",
)
For training pipelines (SFT / GRPO), inference post-selection, and advanced usage, see the companion training repository. For upstream features (vLLM serving, DashScope API, deployment), refer to the upstream Qwen3-TTS repository.
Checkpoint contents
├── model.safetensors # main checkpoint (~3.6 GB)
├── config.json / configuration.json / generation_config.json
├── preprocessor_config.json
├── tokenizer_config.json / vocab.json / merges.txt
└── speech_tokenizer/ # 12 Hz speech tokenizer (~651 MB, same as upstream)
License
Released under the Apache License 2.0, consistent with upstream Qwen3-TTS.
Acknowledgements
Our sincere thanks to the Qwen team for open-sourcing Qwen3-TTS — the base model, tokenizer, and reference training / inference code that make this work possible.