Qwen3-TTS

What is Qwen3-TTS?

Qwen3-TTS is a family of speech generation models released by the Qwen team and published on GitHub, Hugging Face and ModelScope. The 0.6B and 1.7B models turn text into speech in ten languages (Italian, Spanish, Portuguese, Russian, French, German, Korean, Japanese, English and Chinese), along with a few dialect voices. Three model types cover different jobs. CustomVoice offers nine preset voices with instruction-based style control, VoiceDesign creates a new voice from a written description, and the Base models clone a voice from about three seconds of reference audio and can be fine-tuned. A single model handles both streaming and non-streaming output, and the authors report latency as low as 97ms for the first audio packet. You install the qwen-tts Python package, load a model by its Hugging Face name, run a local web demo, serve it with vLLM, or call it through the DashScope API. There is no license fee listed in the README, but you need suitable GPU hardware to run the weights yourself.

Use cases

• Generate speech from text in ten languages
• Clone a voice from a short audio clip
• Design a new voice from a text description
• Control tone, pace and emotion with instructions
• Stream speech for real-time voice apps
• Fine-tune a base voice model
• Serve text-to-speech with vLLM

Who it's for

• Speech and ML engineers
• App developers
• Researchers
• Voice product teams

Why use Qwen3-TTS

• Open weights let you run speech synthesis on your own hardware instead of paying per request to a hosted voice service.
• One model family covers preset voices, voices created from a description and voice cloning, so you can start with the closest fit for your product.
• Streaming output with a reported 97ms first-packet latency suits voice assistants and other real-time apps.
• Instructions in plain language adjust tone, speed and emotion without editing audio afterwards.
• Two sizes, 0.6B and 1.7B, let you trade quality features against memory and speed.

Qwen3-TTS features

• Ten supported languages plus dialect voice profiles
• CustomVoice models with nine preset timbres
• VoiceDesign model that builds a voice from a written description
• Base models that clone a voice from about three seconds of audio
• Natural-language control of timbre, emotion and speaking rate
• Streaming and non-streaming generation from one model
• Qwen3-TTS-Tokenizer-12Hz for encoding and decoding speech
• qwen-tts Python package with batch inference
• Local web UI demo and vLLM usage
• DashScope API access and fine-tuning support

Qwen3-TTS integrations

• Hugging Face
• ModelScope
• vLLM
• DashScope
• FlashAttention

How to get started with Qwen3-TTS

1. Create a fresh Python environment and run pip install -U qwen-tts, optionally adding FlashAttention 2.
2. Load a model such as Qwen3-TTS-12Hz-1.7B-CustomVoice by its Hugging Face name, or from a local download.
3. Call generate_custom_voice with text, language and speaker, or generate_voice_design with a written voice description.
4. For cloning, load a Base model and pass a short reference audio clip with its transcript.
5. Save the returned audio with soundfile, launch the local web demo, or serve the model through vLLM or the DashScope API.

Qwen3-TTS limitations

• Running the models yourself needs compatible GPU hardware, and FlashAttention 2 requires float16 or bfloat16
• The README says other models from the technical report will be released later
• Preset CustomVoice speakers cover a small set of nine timbres
• Instruction control is listed for the 1.7B VoiceDesign and CustomVoice models, not the 0.6B models

Qwen3-TTS pricing

Qwen3-TTS is open source, so you can use and self-host it for free.

Prices can change and may vary by country or exclude tax. Check the official pricing page.

Qwen3-TTS alternatives

The main Qwen3-TTS alternatives listed on Debuly are ElevenLabs, CAMB.AI, Cartesia, Hume AI, Murf AI and AudioCraft. The table compares pricing and platforms; order is A–Z.

1. 

   AudioCraft

   Meta AI research code base for text-to-music (MusicGen), text-to-sound (AudioGen) and EnCodec neural audio compression.

   Audio and voiceDeveloper platforms
   Visit AudioCraft (official site)
2. 

   CAMB.AI

   AI dubbing, text-to-speech, translation and voice cloning across 140+ languages, with a studio app and developer API.

   Audio and voiceVideoDeveloper platforms
   FreemiumVisit CAMB.AI (official site)
3. 

   Cartesia

   Voice AI platform with Sonic text-to-speech, Ink speech-to-text and managed voice agents, priced by credits and minutes.

   Audio and voiceDeveloper platformsAutomation and agents
   FreemiumVisit Cartesia (official site)
4. 

   ElevenLabs

   AI audio platform for text to speech, voice cloning, dubbing, transcription, music and conversational voice agents, via app or API.

   Audio and voiceAutomation and agentsVideoDeveloper platforms
   FreemiumVisit ElevenLabs (official site)
5. 

   Hume AI

   Expressive text-to-speech and empathic voice agent APIs that let developers direct emotion, tone and pacing in generated speech.

   Audio and voiceDeveloper platformsAutomation and agents
   FreemiumVisit Hume AI (official site)
6. 

   Murf AI

   Text-to-speech platform with a voice API, Studio and Dub products, and voice agents, billed per character on the API.

   Audio and voiceDeveloper platformsVideo
   FreemiumVisit Murf AI (official site)

Qwen3-TTS FAQ

Sources & verification

Facts on this page come from the official sources below, each with the date we read it. Written by Debuly; the vendor did not review this page.

Image credits

• Logo: from avatars.githubusercontent.com · fetched 24 Oct 2026 · Vendor trademark, used only to identify the product; removed on request
• Cover: none collected yet, so a generated tile is shown.

Spot an error? Report an issue with this listing. We also remove a logo or image on the owner's request.