Qwen3-TTS adds voice cloning and voice design
Alibaba's Qwen team has released two API-only Qwen3-TTS models: one designs new voices from text descriptions, and the other clones a voice from three seconds of audio across 10 languages.
Alibaba's Qwen team announced two new models in its Qwen3-TTS family: Qwen3-TTS-VD-Flash for voice design and Qwen3-TTS-VC-Flash for voice cloning. Both are available through the Qwen API. As we reported earlier, Qwen3-TTS-Flash added 49 preset voices and Chinese dialects. The new models let users define their own voices or copy existing ones.
Voice design. VD-Flash creates a voice profile from a natural-language description. Instructions can cover timbre, prosody, emotion, persona and background information. Qwen's samples range from short persona tags, such as "Playful Homebody Sis", to long character biographies and attribute-by-attribute specifications covering pitch, speed, volume and accent. Qwen says that on the InstructTTS-Eval benchmark, the model "significantly outperforms" GPT-4o-mini-tts and Mimo-audio-7b-instruct overall. It also says the model beats Gemini-2.5-pro-preview-tts on role-playing tests.
Voice cloning. VC-Flash clones a voice from a 3-second sample. It can then generate speech in that voice in 10 languages: Chinese, English, German, Italian, Portuguese, Spanish, Japanese, Korean, French and Russian. Qwen reports that its average word error rate on the MiniMax TTS Multilingual Test Set is consistently lower than that of MiniMax, ElevenLabs and GPT-4o-Audio-Preview.
Qwen also says both models adjust tone and rhythm to match the meaning of the text and can handle non-standard text formats.
All benchmark results come from Qwen. The announcement does not give pricing, latency figures, model sizes or any open-weight release. It also does not say what safeguards, if any, apply to cloning someone's voice.