# Qwen3-TTS — card for AI agents

Multilingual TTS family (10 languages) with preset voices, voice design from text prompts and 3-second voice cloning.

Category: Voiceover (text-to-speech) · maker Qwen team (Alibaba) · checked 2026-10-25

## Price
- Price: Free
- Free tier: Open weights, free to download.

## License
Apache-2.0 — https://github.com/QwenLM/Qwen3-TTS/blob/main/LICENSE

- commercial: **yes** (Commercial use OK)
  > each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute
  (https://github.com/QwenLM/Qwen3-TTS/blob/main/LICENSE)
- attribution: **required** (Credit required)
  > You must give any other recipients of the Work or Derivative Works a copy of this License
  (https://github.com/QwenLM/Qwen3-TTS/blob/main/LICENSE)
- in app: **yes** (Can ship inside your app)
  > each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute
  (https://github.com/QwenLM/Qwen3-TTS/blob/main/LICENSE)

## Rules to follow
- Apache-2.0 weights and code; commercial use and products allowed.
- When redistributing the weights or code, include the license and keep notices.
- No credit is required inside the generated audio.

## Can an AI agent use it by itself?
**Yes** · package: Agent installs a package or loads it from a CDN

## How to connect
- Package (npm/PyPI)
- PyPI: `pip install qwen-tts`
- Repository: https://github.com/QwenLM/Qwen3-TTS
- Website: https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice

## Notes
One entry for the Qwen3-TTS family: 1.7B CustomVoice, 1.7B VoiceDesign, 1.7B Base (voice clone) and 0.6B variants. Hugging Face API, Oct 25, 2026 (Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice): 1914864 downloads, 2173 likes; GitHub QwenLM/Qwen3-TTS: 14002 stars.

## Numbers (AgentGid data)
- Gid Score: 40
- GitHub stars: 14,511
- Downloads last week: 90,893
- Used in 1 public projects set up for AI coding agents

## Lines to paste into your prompt or AGENTS.md

```
- Qwen3-TTS (Voiceover (text-to-speech)). Multilingual TTS family (10 languages) with preset voices, voice design from text prompts and 3-second voice cloning. Price: free; commercial use OK; credit required.
  Connect: PyPI qwen-tts.
  Rule: Apache-2.0 weights and code; commercial use and products allowed.
  Rule: When redistributing the weights or code, include the license and keep notices.
  Rule: No credit is required inside the generated audio.
  License: https://github.com/QwenLM/Qwen3-TTS/blob/main/LICENSE (checked 2026-10-25; details: https://agentgid.com/assets/qwen3-tts/)
```

## Links
- AgentGid page: https://agentgid.com/assets/qwen3-tts/
- Is it free for commercial use? https://agentgid.com/assets/qwen3-tts/commercial-use/
- Source: https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice
- Source: https://github.com/QwenLM/Qwen3-TTS
- Source: https://github.com/QwenLM/Qwen3-TTS/blob/main/LICENSE
- Source: https://pypi.org/project/qwen-tts/


## Qwen3-TTS: open voices that read your text aloud

Qwen3-TTS is a free, open family of AI voice models from the Qwen team at Alibaba that turns written text into natural speech. It speaks 10 languages, including English, Russian and Chinese. You can pick a ready voice, describe a new voice in words, or copy a voice from a 3-second recording, and run it all on your own computer.

### Where to use it
- Voiceover with a ready voice: Pick one of nine ready voices and get a sound file for a video, a lesson or an app screen. You can add a short note such as "Very happy." to set the mood.
- Describe a new voice in words: Write how the voice should sound, such as age, tone and emotion, and the model makes a matching voice. Handy for game characters or story narrators.
- Speak in a voice from a short clip: Give a few seconds of recorded speech and its words, and new text is read in that voice. Use your own voice or one you have permission to use.
- One character voice for many lines: Design a voice once, save it as a reusable sample, and use it for every line of a game, comic dub or course, so the character always sounds the same.
- Try voices in a local web page: Start a small web page on your computer to type text, pick a voice and listen, before writing any code.
- Read articles aloud in several languages: Add a Listen button to blog posts or help pages and make the audio in the reader's language with the same voice. (idea, not from the official pages)

### How to start
- Make a fresh Python 3.12 environment so its pinned packages do not clash with others.

```
conda create -n qwen3-tts python=3.12 -y
conda activate qwen3-tts
```

- Install the qwen-tts package. The model files download by themselves the first time you load a model by name.

```
pip install -U qwen-tts
```

- Optionally install FlashAttention 2 to use less graphics memory, if your graphics card supports it.

```
pip install -U flash-attn --no-build-isolation
```


### License
Apache 2.0 license: free to use in personal and paid products, and you may change and share the models and code. When you pass the files on, include the license and keep the notices.

### Good to know
- The examples run on an NVIDIA graphics card. Even the small 0.6B model file is about 1.8 GB, so plan disk space and memory.
- Mood notes such as "Very happy." work only with the larger 1.7B models. The small 0.6B voices and the copy-a-voice models do not take them.
- Each ready voice sounds best in its own language, though every voice can speak all ten.
- The page for copying a voice needs a secure connection when opened from another machine, or the browser blocks the microphone.
- A paid hosted version runs on Alibaba Cloud if you do not want to run the models yourself.

Checked 2026-11-10 against: https://github.com/QwenLM/Qwen3-TTS, https://github.com/QwenLM/Qwen3-TTS/blob/main/README.md, https://github.com/QwenLM/Qwen3-TTS/blob/main/LICENSE, https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice, https://huggingface.co/Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice/raw/main/README.md, https://huggingface.co/api/models/Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice/tree/main, https://pypi.org/pypi/qwen-tts/json

### Mistakes to avoid
- Don't: pip install qwen-tts into an existing project environment → Do: create a fresh env first (conda create -n qwen3-tts python=3.12) (The package pins exact versions of transformers and accelerate that can clash with other packages.)
- Don't: attn_implementation="flash_attention_2" with float32 or on unsupported hardware → Do: load with dtype=torch.bfloat16 (or float16) on a supported GPU, or leave out attn_implementation (FlashAttention 2 only works in half precision on compatible hardware.)
- Don't: generate_voice_clone(text=..., ref_audio=clip) without ref_text → Do: pass ref_text with the exact words of the clip, or set x_vector_only_mode=True and accept lower similarity (The clone model expects the transcript of the reference clip.)
- Don't: calling generate_voice_clone many times with the same ref_audio → Do: build the prompt once with create_voice_clone_prompt and pass voice_clone_prompt (Reusing the prompt avoids recomputing the reference features every time.)

### Questions
- Does it speak Russian? Yes. Russian is one of the ten languages, together with English, Chinese, Japanese, Korean, German, French, Portuguese, Spanish and Italian.
- Can I use the voices in a paid app or video? Yes. The models and code are under Apache 2.0, which allows commercial use; keep the license when you pass files on.
- How long a recording do I need to copy a voice? About 3 seconds of speech, together with the words spoken in it.
- Can I copy a voice without typing out the words of the sample? Yes, with x_vector_only_mode=True, but the copy may sound less like the original.

---
AgentGid's catalog is curated: every entry was checked against the maker's official pages (registry, repository, license and pricing pages) on the date shown; a license value shows the quote it stands on wherever AgentGid has one. Usage numbers come from AgentGid's own scan of public repositories set up for AI coding agents. Claims by makers are reported, not tested by AgentGid.
