Qwen3-TTS has become one of the fastest-rising AI searches this autumn. Trend trackers show "qwen3 tts" climbing sharply over the last three months, and it's easy to see why: it's a free, open text-to-speech model family from Alibaba's Qwen team that can design a voice from a written description, speak in preset voices, or clone a voice from a few seconds of audio.
This Qwen3-TTS review explains what the models do, which one to pick, what hardware and setup you need, and where creators should be careful. The capabilities and scores come from Qwen's own model card and announcement. We haven't run our own listening tests, so we don't rank it against paid tools on sound quality.
QUICK ANSWER
Qwen3-TTS is a free, open text-to-speech model family from Alibaba's Qwen team, released as open source on January 26, 2026. It comes in 1.7B and 0.6B sizes and can design a voice from a written description, speak in preset voices, or clone a voice from a few seconds of audio. Clone voices only with permission.
What is Qwen3-TTS?
Qwen3-TTS is "an open-source series of TTS models developed by the Qwen team at Alibaba Cloud", according to the official GitHub repository. Alibaba Cloud announced the open-source release on January 26, 2026 (Alibaba Cloud blog). The family comes in two sizes, 1.7B and 0.6B parameters. Qwen describes the 1.7B models as the best quality and the 0.6B models as a balance of quality and speed.
The Qwen3-TTS models, explained
| Model | What it does |
|---|---|
| 1.7B-VoiceDesign | Creates a new voice from a text description, like "warm, calm male narrator". |
| 1.7B-CustomVoice | Nine built-in voices, with style control through instructions. |
| 0.6B-CustomVoice | The same nine voices in a smaller model, without instruction control. |
| 1.7B-Base / 0.6B-Base | "3-second rapid voice clone" from your audio, and a base for fine-tuning. |
The nine CustomVoice speakers listed on the Hugging Face model card include Ryan and Aiden (native English), Vivian and Serena (Chinese), Ono_Anna (Japanese) and Sohee (Korean). Qwen says each speaker can speak any supported language, but sounds best in its native one.
Which languages does Qwen3-TTS support?
Ten: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian, plus some Chinese dialect voices. Urdu, Hindi and Arabic are not on the list. That's important if you make content for South Asian audiences. For English voiceovers it's a strong option. For Urdu narration, you'll need another tool.
Key features
- Voice from a description: VoiceDesign turns a sentence about age, tone and mood into a voice.
- Emotion and style control: instructions like "Very happy" or "speak in an incredulous tone" change delivery.
- Streaming: Qwen says the first audio packet can arrive after a single character of input, with end-to-end latency as low as 97ms. That matters for live apps and voice agents.
- Long-form audio: Qwen reports word error rates of 2.36% (Chinese) and 2.81% (English) over continuous 10-minute synthesis.
How good is Qwen3-TTS? What Qwen reports
Qwen's announcement says that across its 10-language test set, the voice-clone model reached an average word error rate of 1.835% and a speaker similarity of 0.789, and that it outperformed MiniMax and ElevenLabs on those measures. Those are Qwen's own benchmarks, chosen and run by Qwen. They measure accuracy and similarity, not how pleasant a voice sounds to your audience. Listen to samples yourself before committing.
How to try Qwen3-TTS
There are three routes, from easiest to most technical:
- Hosted API: Alibaba Cloud's Model Studio offers real-time APIs for the custom voice, voice clone and voice design models. Links are on the model card.
- Local web demo: install the
qwen-ttsPython package withpip install -U qwen-tts, then runqwen-tts-demo Qwen/Qwen3-TTS-12Hz-1.7B-CustomVoice --port 8000and open it in your browser. - Python code: load
Qwen3TTSModeland callgenerate_custom_voice,generate_voice_designorgenerate_voice_clone.
Qwen recommends a clean Python 3.12 environment, and its examples run on an NVIDIA GPU (device_map="cuda:0") with bfloat16. FlashAttention 2 is recommended to cut GPU memory use. If you don't have a suitable graphics card, the hosted API is the practical choice.
A simple creator workflow
- Write your script and read it aloud once, to catch awkward lines.
- Use VoiceDesign to describe a narrator voice, or pick a CustomVoice speaker.
- Generate one paragraph at a time, so a mistake doesn't ruin the whole file.
- Listen for mispronounced names and numbers, and spell them out phonetically if needed.
- Edit and mix in your normal audio or video editor.
Voice cloning: the rules you must follow
Voice cloning is where creators get into trouble. Only clone your own voice, or a voice you have clear written permission to use. Cloning someone else to make them "say" things they never said can be deceptive, can break platform rules, and in many places can break the law.
YouTube, for example, says cloning your own voice for voiceovers or dubs doesn't need an AI disclosure, but content that makes a real person appear to say something they didn't does (YouTube Help). Also check the licence on the model card before you use Qwen3-TTS commercially.
Qwen3-TTS: pros and cons
Pros
- Free, open models you can run yourself.
- Voice design from plain-English descriptions.
- Low-latency streaming for real-time apps.
- Strong published accuracy across 10 languages.
Cons
- No Urdu, Hindi or Arabic.
- Local use needs a capable NVIDIA GPU and some Python comfort.
- Benchmarks are self-reported.
Who is Qwen3-TTS for?
It's a good fit for English-language YouTubers, podcasters, app developers building voice features, and students experimenting with AI audio. If you want an all-in-one study helper that can also make audio overviews from your notes, see our Gemini Notebook study guide. If you're choosing a general assistant to help write scripts, our ChatGPT vs Gemini vs Claude comparison will help.
Want to put this into practice? Try our free Text to Speech and Voice Recorder right in your browser, then read YouTube AI Disclosure Rules for Creators (2026) and Apple Creator Studio AI Update on Botvana.
Related: Many open models like this one are shared on Hugging Face. Read why the DOJ is asking about the Nvidia and Hugging Face deal.
Sources and further reading
- QwenLM/Qwen3-TTS repository (GitHub)
- Qwen3-TTS Family is Now Open Sourced: Voice Design, Clone, and Generation! (Alibaba Cloud)
- Qwen/Qwen3-TTS-12Hz-1.7B-Base (Hugging Face)
- Disclosing use of GenAI content - Android phone & tablet (Google Help)
FAQ
Is Qwen3-TTS free?
The models are openly released on GitHub and Hugging Face. Running them locally is free apart from your hardware. The hosted API is paid.
Does Qwen3-TTS support Urdu?
No. It supports 10 languages, and Urdu isn't one of them.
How much audio does voice cloning need?
Qwen describes the Base models as capable of a "3-second rapid voice clone". Giving the transcript of the clip improves quality.
Can I run Qwen3-TTS without a GPU?
Qwen's examples target NVIDIA GPUs. Without one, use the hosted API.
Which Qwen3-TTS model should I start with?
1.7B-CustomVoice for ready-made voices, or 1.7B-VoiceDesign if you want a unique narrator.
Comments
Post a Comment