Qwen3-TTS-Flash adds 49 voices and Chinese dialects
Alibaba's Qwen team has updated its Qwen3-TTS-Flash speech model with more voices, 10 languages and nine Chinese dialects, and claims lower multilingual error rates than MiniMax, ElevenLabs and GPT-4o-Audio-Preview.
The Qwen team announced an update to Qwen3-TTS-Flash, which it calls its flagship text-to-speech model. The model is offered through the Qwen API on DashScope, and demos are hosted on Hugging Face and ModelScope.
The update has three main parts:
- Timbres: The model now has more than 49 preset voices covering a range of genders, ages, regional traits and character types. They have names such as Ryan, Jennifer, Elias, Momo and Vivian.
- Languages and dialects: It supports 10 languages: Chinese, English, German, Italian, Portuguese, Spanish, Japanese, Korean, French and Russian. It can also synthesize Mandarin and eight regional varieties: Hokkien, Wu, Cantonese, Sichuanese, and the Beijing, Nanjing, Tianjin and Shaanxi dialects.
- Prosody: Qwen says the model is much better than the previous version at adjusting speaking rate and prosody to match the input text. It describes the result as close to real human speech.
On the MiniMax TTS multilingual test set, Qwen reports a lower average word error rate (WER) than MiniMax, ElevenLabs and GPT-4o-Audio-Preview. The post does not give the WER figures or results for each language, and it includes no evaluations from other parties.
The announcement includes audio samples in English, Mandarin and the regional dialects. It does not give pricing, latency, model size, or whether the weights will be released. Based on the post, the model is available through the API, with hosted demos for testing.