Independent tool overview

Qwen3-TTS at a glance

Qwen3-TTS is Alibaba's Apache-2.0 speech-generation family, spanning 0.6B and 1.7B local models for preset voices and voice cloning, a 1.7B voice-design model, and hosted streaming APIs.

Visit the official Qwen3-TTS site ↗

Developer: Qwen team at Alibaba Cloud
Open model sizes: 0.6B and 1.7B
Languages: 10
Voice-clone reference: About 3 seconds
Open license: Apache 2.0
Access: Local weights or Model Studio API

Overview

What Qwen3-TTS is

Qwen3-TTS is a family rather than one text-to-speech model. The open release includes 0.6B and 1.7B Base checkpoints for rapid voice cloning and fine-tuning, 0.6B and 1.7B CustomVoice checkpoints with nine built-in speakers, and a 1.7B VoiceDesign checkpoint that creates a voice from a natural-language description.

All released 12Hz checkpoints support streaming and ten languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian. VoiceDesign and the 1.7B CustomVoice model also accept instructions for qualities such as emotion, pace, tone, and prosody. Alibaba reports first-packet synthesis latency as low as 97 milliseconds in its test conditions; real application latency depends on hardware, networking, text buffering, and the serving stack.

Developers can run the Apache-2.0 weights with the qwen-tts Python package, launch a local demo, use vLLM-Omni for offline inference, or call Alibaba Cloud Model Studio. Voice cloning creates an obvious impersonation risk, so production systems should require documented speaker consent, block protected or deceptive use, label synthetic audio where appropriate, and retain an auditable record of the source voice and authorization.

Use cases

Who Qwen3-TTS is best for

The strongest fit depends on the job you need the product to complete, not the size of its feature list.

Self-hosted speech products

Run an Apache-licensed model inside a controlled environment when data locality, customization, or infrastructure ownership matters.

Multilingual narration

Generate speech across ten supported languages using built-in voices or a designed voice, then test each target language with native listeners.

Character and brand voices

Describe a new voice with the 1.7B VoiceDesign checkpoint, then create a reusable clone prompt for consistent lines.

Real-time prototypes

Use streaming-capable checkpoints or the hosted realtime API for assistants and interactive experiences that need early audio output.

Capabilities

Core Qwen3-TTS features

Three model roles

Choose Base for voice cloning and fine-tuning, CustomVoice for nine built-in speakers, or VoiceDesign for a new timbre described in natural language.

0.6B and 1.7B checkpoints

Trade resource use against capability. Base and CustomVoice come in both sizes; the open VoiceDesign checkpoint is 1.7B.

Three-second voice cloning

The Base checkpoints can create a voice prompt from roughly three seconds of reference audio, ideally with an accurate transcript. Embedding-only cloning works without a transcript but may reduce quality.

Natural-language voice design

Describe timbre, persona, age, emotion, pace, pitch, and delivery instead of selecting only from a fixed speaker list.

Nine preset voices

CustomVoice includes Chinese, English, Japanese, Korean, Beijing-dialect, and Sichuan-dialect speaker profiles. Each can speak any supported language, though Qwen recommends the native language for best quality.

Ten-language synthesis

Generate Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian speech, with additional dialect-oriented profiles.

Streaming and local serving

Use one checkpoint for streaming or non-streaming generation, launch the included local web demo, or integrate the qwen-tts Python package directly.

Hosted Model Studio APIs

Call standard or realtime Qwen3-TTS Flash, Instruct, voice-design, and voice-cloning endpoints without managing the model runtime.

Process

How the Qwen3-TTS workflow works

Step 1
Select the model role
Use CustomVoice for a supported preset, VoiceDesign for a novel synthetic persona, or Base only when you have permission to clone a real speaker or need fine-tuning.
Step 2
Benchmark local and hosted routes
Compare the 0.6B and 1.7B checkpoints with Model Studio on pronunciation, expressiveness, first-audio latency, throughput, GPU memory, and cost using representative scripts.
Step 3
Prepare text and references
Normalize numbers and abbreviations, set the target language explicitly when known, and record clean reference audio with an exact transcript for cloning.
Step 4
Document consent
Store who owns the voice, the permitted uses, expiration and revocation terms, source recordings, and approval evidence before generating a clone.
Step 5
Evaluate with native listeners
Test names, acronyms, code-switching, dialects, emotion, long passages, and edge cases with people fluent in every production language.
Step 6
Add release safeguards
Label synthetic audio where appropriate, rate-limit generation, block impersonation abuse, protect reference files, and keep logs that connect every output to an authorized voice profile.

Cost

Qwen3-TTS pricing and free plan

The open Qwen3-TTS code and released weights are available under Apache 2.0 with no per-character license fee, but self-hosters pay for compute, storage, engineering, and monitoring. Alibaba Cloud Model Studio bills hosted Qwen3-TTS by input characters; output audio is not billed. International free quotas below last 90 days after Model Studio activation and are region-specific.

Open checkpoints

$0 model license fee

Download and run the 0.6B or 1.7B checkpoints under Apache 2.0.

Self-hosting compute, bandwidth, storage, and engineering are separate
Available through Hugging Face and ModelScope
qwen-tts Python package and local web demo
Commercial users must still follow applicable law and consent requirements

Qwen3-TTS-Flash API

$0.10 per 10,000 characters

International Singapore-region list price for the standard preset-voice Flash endpoint.

110,000-character international free quota for 90 days on the current model
Input characters are billed; output is free
Realtime Flash is listed at $0.13 per 10,000 characters
Region and dated model version affect availability and pricing

Instruct, Voice Design, or Voice Clone API

$0.115 per 10,000 characters

International list price for the standard Instruct-Flash, VD, and VC endpoints.

110,000-character international free quota for 90 days
Output audio is not billed
Realtime rates are higher and differ by capability
Voice enrollment and voice design creation fees are separate

Hosted voice creation

$0.01 per clone or $0.20 per designed voice

Separate Model Studio charges for creating a reusable cloned or designed voice.

International 90-day quota: 1,000 cloned voices per account
International 90-day quota: 10 designed voices per account
Subsequent synthesis is billed by the selected TTS endpoint
Confirm the deployment region and current quota before launch

Pricing checked August 28, 2026. Check current pricing at the source ↗

Assessment

Qwen3-TTS strengths and limitations

Where it stands out

Apache-2.0 checkpoints provide a permissive path to local and commercial deployment.
The family separates preset-voice, voice-design, and voice-cloning jobs instead of forcing one interface onto every use case.
0.6B checkpoints lower the entry point for constrained local hardware and experimentation.
Ten-language support and instruction-based delivery controls cover a broad set of narration workflows.
The same family supports streaming and batch-style generation.
Hosted APIs give teams an alternative when operating GPU inference is not worthwhile.

What to consider

There is no single best Qwen3-TTS checkpoint: Base, CustomVoice, and VoiceDesign have different capabilities, and the open and hosted model names do not map perfectly one-to-one.
The advertised 97-millisecond latency is a vendor-reported best case, not a guarantee for a particular device, region, or application stack.
Self-hosting still requires a compatible PyTorch environment, substantial model downloads, and enough compute for the selected checkpoint; FlashAttention 2 is optional but hardware- and dtype-dependent.
VoiceDesign is only released as a 1.7B open checkpoint, while the smaller 0.6B release covers Base and CustomVoice.
Preset voices can speak all ten languages, but Qwen recommends each speaker's native language for best quality.
Cloning from a few seconds of audio makes impersonation easy, creating consent, fraud, disclosure, and abuse-management obligations.
Alibaba Cloud prices, free quotas, and model aliases vary by region and version and may change independently of the open weights.

Compare

Qwen3-TTS alternatives

The right alternative depends on the specific output, workflow, controls and budget your project requires.

Miscellaneous

Fish Audio S1

Choose Fish Audio S1 when its speech model, creator-facing workflow, or hosted ecosystem performs better on your target voices and languages.

Explore Fish Audio S1 →

Content Creator

VibeVoice

Choose Microsoft's VibeVoice when long-form, multi-speaker synthesis and its smaller real-time model better fit the workload.

Explore VibeVoice →

Content Creator

ElevenLabs Voice Design V3

Choose ElevenLabs Voice Design when a polished hosted product and prompt-to-voice workflow matter more than self-hosting the model weights.

Explore ElevenLabs Voice Design V3 →

Questions

Qwen3-TTS FAQs

What is Qwen3-TTS?

Qwen3-TTS is Alibaba's speech-generation family for text-to-speech, preset voices, natural-language voice design, and rapid voice cloning. It is available as open 0.6B and 1.7B checkpoints and through Model Studio APIs.

Is Qwen3-TTS open source?

Yes. The official repository and released model cards use the Apache 2.0 license. That removes a per-character model license fee for local use, but not infrastructure cost, consent obligations, or other legal requirements.

Which Qwen3-TTS model should I use?

Use CustomVoice for one of nine built-in speakers, VoiceDesign to describe a new synthetic voice, and Base for cloning an authorized real voice or fine-tuning. Start with 0.6B when lower resource use matters; use 1.7B when quality and control justify more compute.

How much audio is needed to clone a voice?

Qwen describes the Base checkpoints as capable of rapid cloning from roughly three seconds of reference audio. A clean clip and matching transcript generally provide a better basis than embedding-only mode. Only clone a voice with explicit permission.

Which languages does Qwen3-TTS support?

The released 12Hz checkpoints support Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, and Italian. Some preset profiles also target Beijing and Sichuan Chinese dialects.

Can Qwen3-TTS run locally?

Yes. The official qwen-tts Python package loads weights from Hugging Face or a local path and includes a local web demo. Qwen also documents vLLM-Omni offline inference. Hardware needs depend on the model, precision, batch size, and latency target.

How much does the Qwen3-TTS API cost?

For Alibaba Cloud's international Singapore deployment, standard Qwen3-TTS-Flash is currently $0.10 per 10,000 input characters. Instruct, Voice Design, and Voice Clone endpoints are $0.115 per 10,000 characters. Realtime and voice-creation fees differ.

Can Qwen3-TTS stream audio?

Yes. Qwen marks all released 12Hz checkpoints as streaming-capable, and Model Studio offers realtime endpoints. Actual first-audio latency depends on text buffering, model size, hardware, serving software, and network distance.

Bottom line

Our Qwen3-TTS verdict

Qwen3-TTS is one of the more complete open speech families for teams that want a choice between local control and a managed API. The 0.6B and 1.7B sizes, ten languages, three distinct voice workflows, Apache license, and streaming support make it a strong evaluation candidate. The decision should come from a blind listening and operations benchmark, not Alibaba's leaderboard alone, and voice cloning should not ship without consent records, abuse controls, and synthetic-media disclosure rules.

Visit Qwen3-TTS website ↗

AI training for the future of work.

Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.

Join AI University Explore The Rundown University

AI Courses

Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.

Daily Guides

To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.

Workshops

Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.

Community

Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.