Skip to main content

Qwen3 TTS

Alibaba's text-to-speech pair for voiceover and narration. Two variants on the board that trade voice breadth for style control:

  • Qwen3 TTS Flash — 36 expressive voices including 8 Chinese regional dialects (Beijing, Shanghai, Sichuan, Cantonese …).
  • Qwen3 TTS Instruct — 24 standard-Mandarin voices plus a natural-language Instructions field for delivery control.

Capabilities​

FeatureFlashInstruct
Text-to-SpeechYesYes
Voices36 (incl. 8 dialects)24 (standard Mandarin)
Dialect voicesYesNo
Style instructions fieldNoYes
Max Duration180s180s

Which one to pick​

  • Need a native regional accent? → Flash. Pick a dialect voice and the speech comes out with the authentic accent — no prompt engineering required.
  • Need directed delivery? → Instruct. Describe the read in the Instructions field: "newscast pace, authoritative but warm, slow down on numbers".

Instructions Field (Instruct)​

Separate from the script itself, so your text stays clean:

  • Script: 今晚八点,决赛正式打响。
  • Instructions: 体育解说风格,语速快,情绪逐句升高

Pace, tone, emotion, emphasis, and persona directions all work — in Chinese or English.

Prompting Tips​

  • Dialect voices carry the accent by themselves. Write standard text; the voice renders it natively — don't try to spell out accents phonetically.
  • Numbers and units read better with spacing. "3 分 20 秒" reads more reliably than "3'20''".
  • Match voice to content register first, then fine-tune with Instructions (Instruct) or punctuation (Flash).

Limitations​

  • Preset voices only — no reference-audio input
  • Single speaker per run — no dialogue mode (use Gemini 3.1 Flash TTS for two-speaker)
  • 3-minute ceiling per generation

See Also​