Qwen3-TTS
github.com
- Free plan
- Linux
- Web
- API
- Self-hosted
What Qwen3-TTS is
Qwen3-TTS is an open-source series of text-to-speech models from Alibaba Cloud’s Qwen team. It generates speech from text in streaming or non-streaming mode, and supports custom voices, voice design from natural-language descriptions, natural-language voice control, and voice cloning from reference audio. The released collection includes 0.6B and 1.7B parameter models for custom voice or base tasks, plus a 1.7B voice design model. The project lists support for 10 languages and WAV export, and reports synthesis latency as low as 97 ms.
You can install the qwen-tts Python package, download weights from Hugging Face or ModelScope, or run its local Gradio interface with `qwen-tts-demo`. The project recommends a fresh Python 3.12 environment and demonstrates CUDA model loading; FlashAttention 2 is suggested to reduce GPU memory use on compatible hardware. Alibaba Cloud DashScope APIs cover custom voice, cloning, and voice design; API use requires an API key. The repository uses the Apache-2.0 license and permits commercial use. vLLM-Omni supports offline inference, while online serving is planned for later.
Who it suits
Qwen3-TTS suits developers who want open-source speech generation, voice design, or voice cloning and can work with a local Python setup or API. Its Apache-2.0 license permits commercial use.
Good to know, and what to watch
- Supports streaming and non-streaming speech generation.
- Includes voice design and voice cloning.
- Offers 0.6B and 1.7B model variants.
- Local Gradio demo can be launched with qwen-tts-demo.
- Apache-2.0 license permits commercial use.
- Voice cloning uses reference audio and its transcript.
- DashScope API use requires an API key.
- vLLM-Omni currently supports offline inference only.
- CUDA and compatible hardware are relevant to the demonstrated setup.
Plans
Every plan and priceQwen3-TTS open-source models
Open source
Getting started
- Open the Qwen3-TTS GitHub project
- Create a fresh Python 3.12 environment
- Install the qwen-tts package from PyPI
- Download model weights from Hugging Face or ModelScope
- Load a model locally, or launch qwen-tts-demo for the Gradio interface
- For DashScope APIs, configure an API key and install the latest DashScope SDK
Limits to check
The listed open-source plan provides 0.6B and 1.7B models and 10 languages. DashScope Voice Design and Voice Cloning API rates are $0.115 per 10,000 input characters, with a 110,000-character free quota for each and free output.
The Club's review
Qwen3-TTS is an open-source model family for generating speech, designing voices, and cloning voices from reference recordings. It best suits developers and teams that can manage Python and, for local inference, compatible GPU hardware. Its strongest case is flexible voice control with a genuinely free local option; hosted API use is metered and online self-hosted serving is not yet complete.
Overview
Created by Alibaba Cloud’s Qwen team, Qwen3-TTS covers streaming and non-streaming synthesis, custom voices, natural-language voice control, voice design, and voice cloning. The release includes 0.6B and 1.7B models for custom-voice or base tasks, plus a 1.7B voice-design model. Ten languages are supported and audio can be exported as WAV.
The project reports that streaming can deliver its first audio packet after one character, with end-to-end synthesis latency as low as 97 ms. That makes it a plausible fit for interactive applications, although deployed latency will depend on hardware and serving configuration. Cloning requires reference audio and its transcript; using only a speaker embedding can reduce the result’s quality.
Key features
- Multiple generation modes: Streaming is intended for responsive output, while non-streaming generation suits batch rendering and simpler pipelines. Having both avoids forcing every workload into one latency model.
- Voice design: Describe a desired voice in natural language instead of selecting only from preset recordings. This is useful for prototypes, branded characters, and applications needing a particular vocal profile.
- Natural-language control: Prompts can guide the voice during generation, giving developers another control layer beyond text and speaker selection.
- Voice cloning: Reference audio plus a transcript can reproduce a speaker’s characteristics. The transcript is a practical preparation requirement, not an optional extra if consistent quality matters.
- Local web interface: The qwen-tts-demo command launches a Gradio interface locally, giving users a visual way to explore the models without building an application first.
- Model choice and loading: Weights are available through Hugging Face and ModelScope and can be loaded locally or by model ID. The 0.6B option can reduce the footprint relative to 1.7B models, while the dedicated 1.7B design model targets voice creation.
- Developer integration: A PyPI package provides installation, and Alibaba Cloud DashScope real-time APIs cover custom voice, cloning, and design. API use requires an API key and the latest DashScope SDK when using that SDK.
- Inference caveat: vLLM-Omni currently supports offline Qwen3-TTS inference; online serving is planned rather than presently supported. Teams seeking a continuously hosted, production endpoint may therefore need DashScope or another service.
Pricing
Qwen3-TTS open-source models — free. The Apache-2.0 release includes the 0.6B and 1.7B models and supports 10 languages. There is no license charge, making local experimentation and commercial deployment possible without a subscription. The trade-off is operational: you provide the environment, hardware, storage, and maintenance.
For hosted use, Alibaba Cloud lists international Qwen3-TTS Voice Design and Voice Cloning API rates of $0.115 per 10,000 input characters, with output free. Each of those API offerings includes a 110,000-character free quota. This suits applications that prefer managed infrastructure, but usage costs begin after the quota and API credentials and SDK setup are required.
Platforms
Qwen3-TTS is available through API, Linux, self-hosted deployment, and a local web interface. Local installation uses the qwen-tts Python package; the project recommends a fresh Python 3.12 environment. CUDA is demonstrated, and FlashAttention 2 is recommended to reduce GPU memory use when compatible hardware is available. WAV is the stated export format.
Who it's for
Choose Qwen3-TTS if you are a developer, ML engineer, or technical team needing commercial-use speech models, voice cloning, and voice design under an Apache-2.0 license. It is especially appropriate when local control, model downloads, or prompt-driven voice creation matter more than a turnkey hosted dashboard.
It is less suitable for nontechnical users who need a managed application, or for teams requiring online vLLM-Omni serving today. GPU compatibility, Python environment management, transcript preparation for cloning, and the single WAV export format all influence whether it fits a production workflow.
Pros and cons
Pros
- Free, permissively licensed models: Apache-2.0 licensing and commercial use support remove subscription and licensing barriers for self-hosted projects.
- Broad voice control: Design, clone, and steer voices with natural-language controls rather than relying on fixed voices alone.
- Streaming option: The reported 97 ms minimum latency and first-packet behavior are well aligned with responsive interfaces.
- Flexible deployment: Local weights, a Gradio demo, Python packaging, and DashScope APIs cover experimentation through hosted integration.
Cons
- Setup is technical: Python 3.12, CUDA-capable hardware, and optional FlashAttention 2 make local deployment a poor fit for casual users.
- Online self-hosting is incomplete: vLLM-Omni supports offline inference now, with online support still planned.
- Cloning needs prepared inputs: Reference audio must be paired with a transcript for dependable quality.
- Hosted usage is metered: DashScope charges $0.115 per 10,000 input characters after the 110,000-character free quota for each listed API function.
Alternatives
For a broader directory of voice-cloning choices, see Voice Cloning Software; for general speech tools, see Text-to-Speech Software.
Resemble AI is the better choice when you want a paid hosted platform spanning mobile, desktop, web, API, and self-hosted use, with a free plan and trial.
Fish Audio fits users seeking a freemium service with an 8,000-credit monthly free tier and web, desktop, Linux, and API access.
IndexTTS is a straightforward free downloadable model and inference-code option without a hosted plan.
Kits AI is better for a managed freemium workflow with conversion minutes and vocal-generation tools.
Sprag suits buyers who prefer paid hosted speech plans billed around generated usage.
Voicebox offers another free, downloadable voice tool for users comparing self-hosted projects.
VoxCPM is an alternative free and open-source option with no paid tiers or stated usage limits.
Fish Speech provides a free tier with monthly credits, generation and character caps, and public voice slots.
Verdict
Qwen3-TTS is a strong recommendation for technical users who want open-source, commercially usable speech with voice design, cloning, and streaming in one model family. Choose it for local control and flexible prompting; look elsewhere if you need a polished hosted product or online self-hosted serving without managing the underlying infrastructure.
The verdict
Qwen3-TTS offers several speech-generation modes, with local installation and hosted API routes. Consider its setup requirements and offline-serving limitation when choosing how to deploy it.
Questions people ask
Is Qwen3-TTS free?
The open-source models are listed as free and released under Apache-2.0. A free plan is available.
Can it clone voices?
Yes. Voice cloning uses reference audio and its transcript; using only a speaker embedding may reduce quality.
Which platforms are supported?
The listed platforms are API, Linux, self-hosted, and web.
What APIs are available?
Alibaba Cloud DashScope real-time APIs cover custom voice, voice cloning, and voice design. API use requires an API key.
What output format is listed?
WAV is the listed export format.
What the maker publishes
- API access
- Yes source
- API integration
- The project links Alibaba Cloud DashScope real-time APIs for custom voice, voice cloning, and voice design. source
- API pricing
- Alibaba Cloud lists international Qwen3-TTS API rates of $0.115 per 10,000 input characters for Voice Design and Voice Cloning, with output free; the page also lists a 110,000-character free quota for each. source
- API requirements
- The linked Alibaba Cloud real-time synthesis guide says API use requires configuring an API key and installing the latest DashScope SDK when using that SDK. source
- Commercial use
- Yes source
- Download and install
- Models can be downloaded through Hugging Face or ModelScope, and the project provides a `qwen-tts` Python package installable from PyPI. source
- Export formats
- WAV source
- Generation modes
- The models support streaming and non-streaming speech generation, voice design, voice cloning, and natural-language voice control. source
- Hardware and setup
- The README demonstrates model loading on CUDA and recommends FlashAttention 2 to reduce GPU memory use, subject to compatible hardware. source
- Inference integration
- The README states that vLLM-Omni supports offline Qwen3-TTS inference and that online serving is planned for later support. source
- Languages
- 10 source
- Languages
- The project lists support for Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian, as well as multiple dialectal voice profiles. source
- Latency
- The README says the streaming architecture can produce the first audio packet after one character and reports end-to-end synthesis latency as low as 97 ms. source
- License
- The GitHub repository lists the Apache-2.0 license. source
- Local installation
- The project provides a Python package install and says it recommends a fresh Python 3.12 environment. source
- Local web interface
- A local Gradio web UI demo can be launched with the qwen-tts-demo command. source
- Model hosting
- The README links model downloads on Hugging Face and ModelScope and documents loading weights locally or by model ID. source
- Model sizes
- Released models include 0.6B and 1.7B parameter variants for custom voice or base tasks, plus a 1.7B voice design model. source
- Platforms
- web,api,self_hosted source
- Product
- Qwen3-TTS is an open-source series of text-to-speech models developed by the Qwen team at Alibaba Cloud. source
- Serving limitation
- The README says vLLM-Omni supports offline inference currently, with online serving planned for later. source
- Streaming
- The README says the models support streaming and non-streaming generation and reports end-to-end synthesis latency as low as 97 ms. source
- Voice cloning
- Yes source
- Voice cloning
- The Base models can clone a voice from a reference audio clip and are described as capable of rapid cloning from three seconds of input audio. source
Checked against the maker's pages on 7 Oct 2026.