Qwen3-TTS

Verified

Open-source text-to-speech models with voice cloning and voice design

Visit Qwen3-TTS →No rating yetFree

Product preview

Qwen3-TTS

What it is

Qwen3-TTS is an open-source family of text-to-speech models from Alibaba Cloud's Qwen team, released in January 2026 under the Apache-2.0 license. It comes in 0.6B and 1.7B sizes and three types: CustomVoice with nine preset voices and instruction-based style control, VoiceDesign for creating a voice from a written description, and Base for cloning a voice from about three seconds of reference audio. It speaks ten languages, including English, Chinese, Japanese, Spanish and German, and supports streaming with very low latency. Developers install it through pip, download weights from Hugging Face or ModelScope, run a local web demo, or call hosted versions through Alibaba Cloud Model Studio.

Key features

Voice cloning from about 3 seconds of audio
Voice creation from text descriptions
Streaming output with low latency
10 languages plus Chinese dialects

What we liked

+ Permissive Apache-2.0 license
+ Small enough to run locally
+ Hosted API option on Alibaba Cloud

Worth noting

· Needs Python setup and a capable GPU for best speed
· Language list is narrower than some commercial services

Editor's notes

More context from AI BirdView

Open speech models often force a choice between quality and flexibility, and Qwen3-TTS tries to offer both. Three model types cover the most common needs, from ready-made voices to cloning to designing a brand new voice from a sentence of description, and the permissive license means teams can build on it commercially.

Base models clone a voice from roughly three seconds of reference audio and can be fine-tuned.
VoiceDesign creates a voice from a natural-language description, and that voice can then be reused for cloning.
Streaming generation starts audio almost immediately, which the team reports as latency as low as 97 ms.
A pip package, a local web demo and vLLM-Omni support make it practical for developers.
Real-time hosted versions are available through Alibaba Cloud Model Studio for teams that prefer an API.

It is best for developers building voice assistants, narration pipelines or accessibility features who want control over hosting and cost. A practical tip: start with the 0.6B model to prototype on modest hardware, then move to 1.7B when you need voice design or higher expressiveness, and always get consent before cloning anyone's voice.