Multilingual voice agents
Engineering teams that need streaming speech in several supported languages with one provider and controlled voice identity.
Independent tool overview
Voxtral TTS at a glance
Voxtral TTS is Mistral AI's 4B-parameter multilingual text-to-speech model for narration and low-latency voice agents. It supports nine languages, preset or cloned voices, streaming, a $0.016-per-1,000-character API, and noncommercial open-weight deployment.
Visit the official Voxtral TTS site ↗
- Best for
- Developers building multilingual narration, branded voices, localization, and low-latency voice agents
- Model
- Voxtral Mini TTS, a roughly 4B-parameter text-to-speech model
- Languages
- English, French, Spanish, German, Italian, Portuguese, Dutch, Arabic, and Hindi
- Voice adaptation
- Preset voices, reusable saved voices, or one-off reference audio; cloning can work from about 2–3 seconds
- API price
- $0.016 per 1,000 characters, or $16 per 1 million characters
- Latency
- About 0.8 seconds to first PCM audio or about 3 seconds for MP3 in Mistral's documentation
- Open weights
- Available under CC BY-NC 4.0; self-hosting requires a GPU with at least 16GB memory
- Last reviewed
- September 1, 2026
Overview
What Voxtral TTS is
Voxtral TTS turns text into expressive speech in English, French, Spanish, German, Italian, Portuguese, Dutch, Arabic, and Hindi. It can use Mistral's preset voices, a reusable saved voice, or one-off reference audio, and it carries rhythm, accent, intonation, emotion, and even disfluencies from a short voice sample into the result.
The hosted API is aimed at production narration and conversational systems. It supports standard and streaming delivery in MP3, WAV, PCM, FLAC, and Opus. Mistral reports about 90 milliseconds of model processing, but its own documentation puts end-to-end first audio at roughly 0.8 seconds for PCM and three seconds for MP3, a more useful expectation for application design.
Developers can also download the 4B open-weight model, run it with vLLM Omni on a GPU with at least 16GB of memory, and use the bundled reference voices. That release inherits a CC BY-NC 4.0 license, so it is not a free path to commercial self-hosting. Commercial teams should use the hosted API or negotiate appropriate rights rather than assuming open weights mean unrestricted use.
Voxtral can clone a voice from only a few seconds of audio. Mistral's API documentation requires explicit consent and prohibits impersonation, fraud, deception, misinformation, and privacy-invasive content, but the documented creation call accepts a base64 audio sample without describing a live speaker-verification step. Teams must build durable consent, disclosure, access controls, and misuse monitoring around the model.
Use cases
Who Voxtral TTS is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Consented brand narration
Companies with explicit speaker rights that want a reusable voice for support, training, podcasts, or product experiences.
Cross-language localization
Teams testing whether a consented voice can remain recognizable across another supported language, with native-speaker review.
Noncommercial self-hosted research
Researchers and experimenters with suitable GPU infrastructure who can comply with the open model's CC BY-NC 4.0 license.
Capabilities
Core Voxtral TTS features
Nine-language synthesis
Generates speech in nine documented languages and supports code-mixing and cross-lingual voice transfer.
Zero-shot voice cloning
Adapts to a speaker from a short audio prompt, including accent, pacing, intonation, and emotional style.
Saved voice profiles
Stores a reference sample as a reusable voice ID so applications do not resend the audio with every request.
One-off reference audio
Generates speech directly from a temporary reference clip when a persistent voice profile is unnecessary.
Streaming output
PCM and other output formats support interactive voice experiences, with PCM providing the fastest documented first audio.
Multiple audio formats
Returns MP3, WAV, PCM, FLAC, or Opus for different quality, storage, and streaming requirements.
Open-weight deployment
The 4B model can be self-hosted with vLLM Omni on suitable hardware for noncommercial uses allowed by its license.
Content moderation
The hosted TTS API can reject text that violates acceptable-use controls.
Process
How the Voxtral TTS workflow works
Step 1
Choose hosted or self-hosted deployment
Use the hosted API for commercial production and managed scaling; use open weights only when the noncommercial license and infrastructure fit.
Step 2
Document voice rights
Obtain explicit written consent covering cloning, storage, languages, scripts, distribution, commercial contexts, duration, and withdrawal.
Step 3
Select or record a voice
Prefer a preset or designed voice when a real person's identity is unnecessary. For cloning, record only the consenting speaker in a controlled setting.
Step 4
Create the profile securely
Save the sample as a restricted voice ID or provide one-off reference audio, and limit both paths to authorized systems and staff.
Step 5
Normalize the script
Spell out ambiguous numbers and abbreviations, remove rich formatting and emojis, and keep segments under 300 words for best results.
Step 6
Choose the delivery format
Use PCM when first-audio latency matters, MP3 for broad compatibility, or lossless formats for production editing.
Step 7
Review every target language
Have a native or proficient reviewer check names, numbers, meaning, politeness, pronunciation, accent, and cultural fit.
Step 8
Disclose and escalate
Tell listeners the voice is AI-generated, identify the responsible organization, and provide a human path for consequential interactions.
Step 9
Monitor, revoke, and delete
Log usage, block unsafe scripts, respond to misuse, honor consent withdrawal, remove the stored voice, and retire exported audio where possible.
Cost
Voxtral TTS pricing and free plan
Voxtral TTS is priced by input characters through Mistral's hosted API. The downloadable model has no API fee but is noncommercial under CC BY-NC 4.0 and requires self-managed GPU infrastructure. General Mistral account credits and enterprise arrangements may offset or change the effective cost.
Mistral Free
$0 account plan
Mistral's current general Free plan includes limited Studio access and $10 per month in API credits.
- Credits are shared across eligible API usage rather than a dedicated unlimited TTS tier
- Account and model availability can vary
- Verify rate limits and billing activation in the live console
Hosted Voxtral TTS API
$0.016 per 1K characters
Usage-based access to voxtral-mini-tts-latest through Mistral's speech endpoint.
- Equivalent to $16 per 1 million input characters
- Voice storage, application infrastructure, telephony, upstream model, and monitoring costs may be separate
- Test retries and regeneration because they also consume usage
Open weights
No API fee; infrastructure required
Downloadable Voxtral-4B-TTS-2603 weights for uses permitted by CC BY-NC 4.0.
- Noncommercial license; attribution and license compliance required
- At least 16GB of GPU memory is documented for serving
- You operate security, scaling, moderation, monitoring, updates, and storage
Enterprise
Contact sales
Commercial private deployments, custom models, support, and governance require a negotiated arrangement.
- Clarify commercial self-hosting rights in writing
- Confirm data region, retention, service levels, incident response, and voice-consent controls
- Obtain legal review for regulated and public-facing deployments
Pricing checked . Check current pricing at the source ↗
Assessment
Voxtral TTS strengths and limitations
Where it stands out
- One model covers nine widely used languages and supports cross-lingual voice transfer
- Short reference audio makes rapid prototyping and voice adaptation easy
- Saved voice IDs and one-off reference clips support both persistent and temporary workflows
- Streaming PCM offers substantially faster first audio than compressed MP3 in Mistral's published measurements
- The hosted price is simple to model from character volume
- Open weights give researchers control over deployment and data flow when the noncommercial license fits
- Multiple output formats cover live agents, editing, archiving, and bandwidth-sensitive delivery
- Mistral publishes explicit consent, disclosure, impersonation, fraud, and harmful-content restrictions in the voice documentation
What to consider
- A two- or three-second clip lowers the technical barrier to cloning and raises the risk of nonconsensual use
- The documented voice-creation API accepts an audio sample but does not describe a live speaker-presence or ownership-verification check
- Policy language does not replace technical controls, written consent, review, access logging, and rapid revocation
- Do not clone any person without explicit, use-specific permission covering the actual scripts, languages, channels, and commercial context
- Generated speech must not be used for deceptive impersonation, fake endorsements, fraud, misinformation, harassment, defamation, or fabricated evidence
- Listeners should be told when audio is synthetic wherever they could reasonably believe the speaker recorded or approved the exact words
- A cloned voice must never serve as identity authentication or proof that a person authorized a payment, contract, vote, prescription, or safety instruction
- Voice, biometric, publicity, labor, election, recording, consumer-protection, and telemarketing rules vary by jurisdiction
- AI-generated or cloned voices used in U.S. calls can trigger FCC artificial or prerecorded voice consent and identification requirements
- The nine supported languages do not cover many global markets, and code-mixing or cross-language output may still mispronounce names or alter meaning
- Cross-lingual synthesis can falsely imply that the source speaker personally speaks or endorses a message in another language
- Mistral's quality comparison with ElevenLabs was vendor-run and used three annotators per pair; it should not be treated as an independent universal benchmark
- Model processing latency is not end-to-end application latency: Mistral documents about 0.8 seconds to first PCM audio and about three seconds for MP3
- Prompts longer than 300 words may reduce quality even though the API can interleave arbitrarily long generations
- Numbers, symbols, abbreviations, markdown, emojis, names, and domain-specific terms require script normalization and testing
- The model can mirror unwanted disfluencies, accents, background artifacts, or emotional cues from the reference
- Content moderation can reject requests and should not be the only abuse prevention layer
- The open-weight release is CC BY-NC 4.0, so commercial self-hosting is not permitted merely because the files are downloadable
- Self-hosting requires suitable GPU hardware plus engineering for scaling, security, observability, moderation, upgrades, and incident response
- A TTS model only speaks the supplied script; if an upstream language model or business system is wrong, the natural voice can make the error more persuasive
- Financial updates, identity verification, industrial instructions, vehicle alerts, healthcare, public-service notices, and emergency content require authoritative data, deterministic controls, human escalation, and domain review
- Deleting a saved voice cannot recall audio already downloaded, cached, syndicated, or published elsewhere
Compare
Voxtral TTS alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Miscellaneous
xAI Custom Voices
Choose xAI Custom Voices for a hosted TTS and real-time voice stack whose console includes a documented live passphrase and speaker-similarity check.
Explore xAI Custom Voices →Marketing
ElevenLabs
Choose ElevenLabs for a mature end-to-end voice platform with creator tools, dubbing, agents, and multiple cloning workflows.
Explore ElevenLabs →Content Creator
Qwen3-TTS
Choose Qwen3-TTS when open-model experimentation and a different licensing or deployment tradeoff matter more than Mistral's hosted API.
Explore Qwen3-TTS →Questions
Voxtral TTS FAQs
What is Voxtral TTS?
Voxtral TTS is Mistral AI's 4B-parameter multilingual speech-generation model for narration, cloned voices, localization, and low-latency voice agents.
What languages does Voxtral TTS support?
English, French, Spanish, German, Italian, Portuguese, Dutch, Arabic, and Hindi. It also supports code-mixing and cross-lingual voice transfer.
How much does the Voxtral TTS API cost?
As of September 1, 2026, Mistral lists the hosted model at $0.016 per 1,000 input characters, equivalent to $16 per 1 million characters.
Is Voxtral TTS open source?
It is open weight, not unrestricted open source. The downloadable model inherits CC BY-NC 4.0, which requires attribution and excludes commercial use.
Can I self-host Voxtral TTS commercially?
Not under the published CC BY-NC 4.0 weights alone. Commercial self-hosting requires separate rights or an enterprise agreement with Mistral.
What hardware does self-hosted Voxtral need?
Mistral's model card says the BF16 model can run on a single GPU with at least 16GB of memory using current vLLM and vLLM Omni versions.
How much audio is needed to clone a voice?
Mistral says adaptation can work from about two to three seconds of audio. A longer, clean, representative sample should be tested because very short clips cannot capture every context.
Does Mistral verify that I own a cloned voice?
The documentation requires explicit consent and prohibits impersonation, but the published API creation example accepts a base64 sample without describing a live ownership check. Build and document your own consent verification.
Can Voxtral use a voice without saving it?
Yes. Speech generation can take one-off reference audio, or you can create a reusable saved voice and pass its voice ID.
How fast is Voxtral TTS?
Mistral reports about 90 milliseconds of model processing, but end-to-end first audio is approximately 0.8 seconds for PCM and three seconds for MP3. Network, queueing, script processing, and application logic add latency.
Can Voxtral generate long audio?
The model natively generates up to about two minutes, while Mistral says its API can handle longer output through interleaving. The documentation recommends keeping text segments under 300 words for quality.
Does Voxtral support streaming?
Yes. PCM is the recommended low-latency format; MP3, WAV, FLAC, and Opus are also supported.
Do I need permission to clone a voice?
Yes. Mistral explicitly requires consent. Use a written agreement that covers cloning, storage, commercial use, languages, channels, generated scripts, distribution, duration, and withdrawal.
Should Voxtral audio be labeled as AI-generated?
Yes whenever a listener could believe the person recorded or approved the exact message, and wherever law or platform rules require it.
Bottom line
Our Voxtral TTS verdict
Voxtral TTS is a strong technical option for multilingual voice applications: nine languages, short-reference adaptation, streaming, straightforward API pricing, and a workable noncommercial self-hosting path. Its ease of cloning demands stronger controls than the API documents technically enforce. Use a preset or designed voice when identity is unnecessary; otherwise require written consent, visible disclosure, locked-down voice assets, native-language review, and human control over every consequential workflow.
Visit Voxtral TTS website ↗Similar Business Operations Tools
View all Business Operations AI tools →
Claude Marketplace
Claude Marketplace - Anthropic's enterprise hub for buying Claude-powered partner tools
Gemini 3.1 Flash Live
Gemini 3.1 Flash Live - Google's low-latency voice model for building real-time AI agents
GLM-OCR
GLM-OCR - Z AI's SOTA multimodal model for extracting text and images from documents
Lovable Payments
Lovable Payments - Lovable's one-conversation setup for adding payments, tax, and multi-currency support to AI-built apps
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.