Self-hosted narration
Teams generating multilingual product narration, explainers or accessibility audio with an open model and a fixed voice roster.
Independent tool overview
Qwen3-TTS CustomVoice 1.7B at a glance
Qwen3-TTS CustomVoice 1.7B is an Apache-licensed text-to-speech model that speaks ten languages through nine built-in voices, with streaming output and natural-language control over delivery.
Visit the official Qwen3-TTS CustomVoice 1.7B site ↗
- Product type
- Open-weight multilingual text-to-speech model
- Model size
- 1.7 billion parameters
- Built-in speakers
- Nine preset voices
- Languages
- Ten documented languages
- Output modes
- Streaming and non-streaming
- License
- Apache 2.0
- Pricing checked
- August 31, 2026
Overview
What Qwen3-TTS CustomVoice 1.7B is
Qwen3-TTS CustomVoice 1.7B is the preset-speaker member of Qwen's open speech-generation family. It turns text into speech using nine named voices and can follow instructions about emotion, pace, tone and prosody.
The model supports Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. Qwen recommends using each speaker's native language for the best quality, although every listed speaker can generate any supported language.
Despite the original directory description, this CustomVoice checkpoint does not clone a person's voice from a three-second recording. Rapid voice cloning belongs to the separate Qwen3-TTS Base checkpoints; VoiceDesign is another separate model. That distinction matters for both product selection and consent controls.
Use cases
Who Qwen3-TTS CustomVoice 1.7B is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Expressive scripted speech
Dialogue and character lines that benefit from instructions for emotion, rhythm, speaking rate and delivery.
Low-latency prototypes
Developers testing streaming speech for assistants or interactive experiences before selecting production infrastructure.
Capabilities
Core Qwen3-TTS CustomVoice 1.7B features
Nine preset timbres
The official roster includes Vivian, Serena, Uncle_Fu, Dylan, Eric, Ryan, Aiden, Ono_Anna and Sohee.
Ten-language synthesis
Generates Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian.
Instruction-based delivery
A free-text instruction can guide emotion, tone, rate and prosody for the 1.7B checkpoint.
Streaming architecture
The family supports streaming and non-streaming generation; Qwen reports end-to-end first-packet latency as low as 97 ms under its test conditions.
Batch generation
The Python interface accepts lists of text, language, speaker and instruction values for multiple outputs.
Local Python package and demo
The qwen-tts package loads checkpoints directly and includes a local web demo command.
Open model files
Weights are available through Hugging Face and ModelScope under Apache 2.0, with vLLM-Omni offline-inference examples.
Process
How the Qwen3-TTS CustomVoice 1.7B workflow works
Step 1
Choose the correct Qwen3-TTS checkpoint
Use CustomVoice for the nine preset speakers, VoiceDesign for a described synthetic persona, or Base for consented reference-audio cloning.
Step 2
Install in an isolated environment
Use the current qwen-tts package and a compatible GPU stack; Qwen recommends FlashAttention 2 to reduce memory use where supported.
Step 3
Select language, speaker and direction
Set the known target language explicitly, choose a speaker and write a concise performance instruction without putting secrets in the text.
Step 4
Generate short review batches
Test pronunciation, names, numbers, code-switching, emotion, noise, latency and consistency across the actual scripts and target devices.
Step 5
Approve and disclose the audio
Have a fluent reviewer listen to the full output, correct errors, document the synthetic voice and obtain the rights needed for script, music and distribution.
Cost
Qwen3-TTS CustomVoice 1.7B pricing and free plan
The CustomVoice checkpoint is free under Apache 2.0, but self-hosting requires suitable compute. Alibaba Cloud also sells related managed Qwen3-TTS speech APIs; those model IDs are hosted services and should not be assumed to be the identical 1.7B checkpoint.
Open model
Free under Apache 2.0
Download and operate the 1.7B CustomVoice weights.
- No model license fee
- Hardware, storage and engineering are not included
- Use of voices and generated content remains subject to applicable rights and law
Self-hosted inference
Compute-dependent
Run locally or on rented accelerators using qwen-tts or a compatible runtime.
- Cost varies by GPU, traffic and latency target
- Streaming production requires your own serving and monitoring
- Benchmark quality and real-time factor on target hardware
Alibaba Cloud Qwen3-TTS Instruct API
$0.115 per 10,000 characters
International Singapore list price for the related qwen3-tts-instruct-flash hosted model.
- Output is free under the listed billing rule
- 110,000-character introductory quota under documented conditions
- Real-time Instruct pricing is $0.143 per 10,000 characters
- Hosted model IDs are not labeled as the open 1.7B checkpoint
Pricing checked . Check current pricing at the source ↗
Assessment
Qwen3-TTS CustomVoice 1.7B strengths and limitations
Where it stands out
- Open weights and Apache 2.0 licensing support private, controlled deployments.
- Ten documented languages cover several major narration markets.
- Nine distinct preset speakers simplify repeatable voice selection.
- Instruction control adds emotion and pacing without training a custom voice.
- Streaming and batch interfaces cover interactive and offline workflows.
- The official package, demo and model cards provide a relatively direct setup path.
What to consider
- CustomVoice is not the three-second voice-cloning model; reference-audio cloning requires a separate Base checkpoint.
- Only nine preset speakers are included, and voice identity cannot be freely specified the way VoiceDesign allows.
- Qwen recommends a speaker's native language for best quality, so cross-language output may be less natural or consistent.
- The reported 97 ms latency is a vendor result under particular conditions, not a promise for every device or deployment.
- Names, numbers, abbreviations, mixed-language text and domain terminology can be mispronounced.
- Emotional instructions can produce inconsistent intensity, pacing or unintended performance choices.
- Synthetic speech can be misused for impersonation, fraud or deceptive media; applications need authentication, consent, disclosure and abuse monitoring.
- High-impact uses such as emergency, medical, legal, financial or identity verification require human review and must not rely on generated speech alone.
Compare
Qwen3-TTS CustomVoice 1.7B alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Marketing
ElevenLabs
A managed commercial speech platform with voice libraries, cloning and production tooling for teams that do not want to self-host.
Explore ElevenLabs →Content Creator
VibeVoice
An open Microsoft speech model oriented toward long-form and multi-speaker generation.
Explore VibeVoice →Questions
Qwen3-TTS CustomVoice 1.7B FAQs
What is Qwen3-TTS CustomVoice 1.7B?
It is an open Qwen text-to-speech checkpoint with nine preset speakers, ten languages, streaming generation and instruction control over delivery.
Can this model clone a voice from three seconds of audio?
No. The three-second rapid-cloning capability belongs to Qwen3-TTS Base checkpoints. CustomVoice uses a fixed list of preset speakers.
Which languages does it support?
Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. Qwen recommends each preset speaker's native language for the strongest quality.
Which voices are included?
Vivian, Serena, Uncle_Fu, Dylan, Eric, Ryan, Aiden, Ono_Anna and Sohee. They cover Chinese and regional Chinese profiles plus native English, Japanese and Korean voices.
Is Qwen3-TTS CustomVoice open source?
The model weights and repository are published under Apache 2.0. You still pay for infrastructure and must comply with content, voice, privacy and distribution rights.
Can it run locally?
Yes, on compatible hardware. Qwen documents its qwen-tts Python package and local demo, recommends FlashAttention 2 where supported, and provides vLLM-Omni offline examples.
How much does it cost?
The open checkpoint has no model license fee. Self-hosting costs depend on hardware and traffic. A related Alibaba Cloud Instruct API is listed at $0.115 per 10,000 characters in Singapore, but it is not identified as the identical 1.7B checkpoint.
What safeguards should a speech app add?
Obtain rights to scripts and voices, disclose synthetic audio, block impersonation and fraud, protect logs and credentials, watermark or retain provenance where appropriate, and require human approval for sensitive messages.
Bottom line
Our Qwen3-TTS CustomVoice 1.7B verdict
Qwen3-TTS CustomVoice 1.7B is a practical open option when nine stable preset speakers and multilingual instruction control are enough. Its biggest documentation trap is the product boundary: it does not perform the family's three-second voice cloning, so teams needing a copied or newly designed identity must choose another checkpoint and apply stricter consent controls.
Visit Qwen3-TTS CustomVoice 1.7B website ↗Similar Consumer Tools
View all Consumer AI tools →
Tiny Aya
Tiny Aya - Cohere Labs' new open-source multilingual small model covering 70+ languages in just 3.35B parameters
Claude Sonnet 4.6
Claude Sonnet 4.6 - Anthropic's upgraded mid-tier model rivaling Opus at 1/5 the cost with 1M-token context
Claude Opus 4.6
Claude Opus 4.6 - Anthropic's upgrade to its most powerful model line, featuring multi-agent collaboration, 1M context window, and new Office integrations
Lyria 3
Lyria 3 - Google's new AI music generation model for creating 30-second, customizable audio clips from text, image, or video prompts.
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.
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.