Qwen-TTS Now Speaks Chinese Dialects Too

The latest version of Qwen-TTS, qwen-tts-latest (or qwen-tts-2025-05-22), has been released. Trained on large-scale speech data, it aims for human-like naturalness and expressiveness, automatically adjusting prosody, speed, and emotional intonation to match the input text.

The most notable change is Chinese dialect support. It can currently generate 3 dialects: Pekingese, Shanghainese, and Sichuanese.

A total of 7 voices are supported.

• Chinese-English bilingual voices: Cherry, Ethan, Chelsie, Serena
• Dialect voices: Dylan (Pekingese), Jada (Shanghainese), Sunny (Sichuanese)

On the performance side, results based on SeedTTS-Eval have also been presented. WER and SIM were disclosed for Chelsie, Serena, Ethan, and Cherry; for example, Cherry recorded zh WER 1.209, en WER 1.967, hard WER 6.069, along with zh SIM 0.799, en SIM 0.664, and hard SIM 0.801.

Usage is simple. You call dashscope.audio.qwen_tts.SpeechSynthesizer.call() from the Qwen API to synthesize speech, then download the returned audio URL using requests. The example code reads the DASHSCOPE_API_KEY environment variable, synthesizes with voice="Dylan" and model="qwen-tts-latest", and saves it as a WAV file.

The key point is that Qwen-TTS has expanded beyond simple TTS into an expressive speech model that encompasses both Chinese-English bilingual speech and Chinese dialects. More languages and style options are planned to be added in the future.