Open-weight text-to-speech
Voxtral TTS
Voxtral TTS turns text into natural speech across 9 languages, supports 2-3 seconds of reference audio for zero-shot voice cloning, and serves both streaming and batch experiences with low latency.
Why Voxtral TTS
Open weights, expressive prosody, and deployment paths that fit real production constraints.
- Zero-shot voice cloning from short 2-3 second audio prompts
- Natural prosody, emotion transfer, and voice-as-instruction behavior
- 24 kHz output with streaming and batch inference for agent workflows
Live workspace
Voxtral TTS Demo
Explore the Voxtral TTS demo below.
Turn text into natural speech and create voices for your next project.
Loading Voxtral TTS demo…
Please wait while the demo prepares.
Benefits
Why teams evaluate Voxtral TTS
Voxtral TTS combines open deployment options with speech quality that feels natural in user-facing products.
Natural speech and emotion
Generate realistic, expressive speech that preserves rhythm, intonation, and emotional color instead of flat robotic narration.
Fast voice adaptation
Clone and adapt voices from short prompts, making it easier to prototype branded assistants and personalized narrators.
Low-latency voice delivery
Use streaming and batch generation paths for voice agents, live experiences, and high-throughput speech pipelines.
Use cases
Where Voxtral TTS fits best
The model is positioned for production voice applications where quality, control, and deployment flexibility matter.
Customer Support
Build support assistants with natural phone-ready speech, voice consistency, and responsive streaming outputs.
Financial Services
Power KYC flows, onboarding agents, and transaction guidance with branded voices and auditable outputs.
Real-time Translation
Pair multilingual speech generation with translation pipelines to deliver fluent spoken experiences in multiple regions.
Automotive Systems
Deploy in-vehicle prompts, copilots, and infotainment voices that feel more human without proprietary lock-in.
Features
Core Voxtral TTS capabilities
Official model and docs highlights condensed into the capabilities most relevant for product and developer teams.
9-language support
English, French, Spanish, Portuguese, Italian, Dutch, German, Hindi, and Arabic with code-mixing and cross-lingual cloning support.
20 preset voices
Start from built-in voices for rapid experimentation before saving reusable profiles in production workflows.
2-3 second voice cloning
Use a short audio sample to transfer voice identity, style, accent, and emotional rendering without retraining.
Streaming and batch inference
Choose low-latency streaming for interactive systems or batch generation for larger content pipelines.
24 kHz audio output
Export speech in WAV, PCM, FLAC, MP3, AAC, and Opus formats for playback, post-processing, or API delivery.
Open weights with vLLM Omni
Serve locally through vLLM Omni and run the 4B model on a single GPU with at least 16 GB of memory.
Build a listening test
Choose a small set of sentences that reflects your actual audience. Include a greeting, a practical instruction, a question, and a sentence containing a name or number. For multilingual work, ask a fluent speaker to review both the written text and the generated audio. A translated sentence can be grammatically correct while sounding unnatural when spoken.
Compare voice and delivery separately
Keep the script unchanged while comparing voices, then keep the chosen voice unchanged while editing the script. This helps you distinguish a voice preference from a pronunciation or pacing problem. When using reference audio, choose material you are authorized to use and listen for background noise, overlapping speakers, or abrupt cuts before evaluating the result.
For an assistant, test the first audible response and the complete answer as separate parts of the experience. A quick start is useful only if the rest of the message remains understandable. For narration, listen across paragraph boundaries and check whether the speaker still sounds consistent after a change in topic.
Match the delivery channel
Review a sample through the device or application where it will be heard. A clip that sounds clear through headphones may behave differently through a phone speaker or a conferencing system. Check important words at a comfortable volume and avoid comparing two candidates at noticeably different playback levels.
Save the input text, voice choice, and relevant settings with each accepted sample. Revisit that small reference set when changing your integration. Keep the full written answer available when audio cannot play, and give listeners a way to stop it. These checks help you evaluate a voice workflow without treating model processing figures as a guarantee for your entire application.
The demo and linked products are separate experiences. If the demo cannot load, follow the official model and documentation links for the project's supported setup, or inspect the alternative service before deciding whether its workflow suits your needs.
Make multilingual reviews comparable
Use a short script with the same purpose in every language, such as a welcome message followed by an instruction. Let each reviewer adapt the wording naturally instead of forcing the same sentence structure. Ask reviewers to note unclear words and misplaced emphasis separately from their personal preference for a voice. Keep those notes next to the corresponding audio. This makes it easier to distinguish a translation issue from a speech issue and to decide whether a revision improved the intended message.
FAQ
Voxtral TTS FAQ
Quick answers for teams exploring Voxtral TTS for production voice applications.
What license does Voxtral TTS use?
The official Hugging Face model card lists Voxtral TTS under a CC BY-NC 4.0 family license because the bundled reference voices inherit that licensing. Review the model card before commercial deployment.
Which languages does Voxtral TTS support?
Official docs list English, French, Spanish, Portuguese, Italian, Dutch, German, Hindi, and Arabic, with support for cross-lingual voice cloning.
How much reference audio do I need for voice cloning?
Mistral documents zero-shot voice cloning from as little as 2-3 seconds of audio, which is enough to transfer voice identity and style.
Does Voxtral TTS support streaming?
Yes. Mistral documents low-latency streaming with roughly 90 ms model processing time, alongside batch generation for offline workflows.
How do I self-host Voxtral TTS?
The model card recommends vLLM Omni for serving. The 4B model can run on a single GPU with at least 16 GB memory, depending on your deployment configuration.
Where can I try Voxtral TTS right now?
You can test the official Hugging Face Space embedded above, inspect the model card on Hugging Face, and read the API docs in Mistral's text-to-speech documentation.