Qwen3-TTS-TBLive-CustomVoice
A CustomVoice checkpoint fine-tuned from TaoLiveAIGC/Qwen3-TTS-TBLive-Base. It ships six built-in open-source speakers spanning three live-streaming styles — professional presentation (专业讲解), warm recommendation (贴心推荐), and energetic promotion (激情促销) — each in a female and a male voice, plus inline instruction-control tokens for pause and speaking-rate control.
Built-in speakers
Pass the speaker ID to generate_custom_voice(text, speaker=...):
| Speaker | Gender | Style | Speaker ID |
|---|---|---|---|
| 专业讲解女 | Female | Professional presentation (专业讲解) | 115 |
| 专业讲解男 | Male | Professional presentation (专业讲解) | 96 |
| 温柔女音 | Female | Warm recommendation (贴心推荐) | 26 |
| 温柔男音 | Male | Warm recommendation (贴心推荐) | 80 |
| 激情促销女 | Female | Energetic promotion (激情促销) | 4 |
| 激情促销男 | Male | Energetic promotion (激情促销) | 1 |
Instruction control tokens
Insert the following tokens directly into the input text; they are interpreted inline at the position where they appear.
Silence / pause tokens — each token synthesizes a pause within a fixed duration band:
| Token | Pause duration |
|---|---|
<sil_L2> | 0.55 – 0.70 s |
<sil_L3> | 0.80 – 0.95 s |
<sil_L4> | 1.05 – 1.20 s |
<sil_L5> | 1.30 – 1.45 s |
<sil_L6> | 1.55 – 2.50 s |
Speed tokens — like silence tokens, they can be inserted anywhere in the text and adjust the speaking rate of the speech that follows:
| Token | Effect |
|---|---|
[speed_e2] | Faster speech |
[speed_d2] | Slower speech |
Quickstart
pip install -U qwen-tts
import torch
from qwen_tts import Qwen3TTSModel
model = Qwen3TTSModel.from_pretrained(
"TaoLiveAIGC/Qwen3-TTS-TBLive-CustomVoice",
device_map="cuda:0",
dtype=torch.bfloat16,
)
wavs, sr = model.generate_custom_voice(
text="[speed_e2] 家人们，最后一波福利来了！<sil_L3> 库存不多，喜欢的抓紧下单！",
speaker="4", # 激情促销女
)
For speaker registration (baking new speakers into a fine-tuned checkpoint), training pipelines, and inference post-selection, 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, six speakers baked in)
├── config.json / configuration.json / generation_config.json
├── preprocessor_config.json
├── tokenizer_config.json / vocab.json / merges.txt
├── tokenizer.json / added_tokens.json / special_tokens_map.json
└── 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.