What it does
Qwen3-TTS is a series of open-source text-to-speech models from the Qwen team at Alibaba Cloud, released under the Apache-2.0 license. The released checkpoints come in 0.6B and 1.7B sizes, alongside a speech tokenizer that encodes audio to codes and decodes it back. The variants cover different jobs. A VoiceDesign model builds a voice from a written description, CustomVoice models offer nine preset voices whose style is steered by instructions, and Base models clone a voice from a roughly three-second audio sample and can also be fine-tuned. The models cover ten languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. The repository describes streaming generation with end-to-end latency as low as 97ms, and control over tone, speaking rate and emotion through natural-language instructions. The models can be run through the qwen-tts Python package, a local web demo, or vLLM-Omni (offline inference only for now). Alibaba Cloud also offers real-time DashScope API options for the custom voice, voice clone and voice design models.
Capabilities
- Voice design from a written description
- Voice cloning from a short reference clip
- Nine preset voices with instruction-based style control
- Streaming generation with latency as low as 97ms
- Ten languages plus dialect voice profiles
- Fine-tuning instructions included
When to use it
- Adding spoken output to an app in one of ten supported languages
- Creating a custom voice from a text description or a short sample
- Running a speech model on your own hardware or through the DashScope API
- Building real-time voice features that need a fast first audio packet
