Key Takeaways
- Voxtral 3 is Mistral AI's newest voice model generation, announced on October 20, 2026, and it ships as a pair: Transcribe for speech-to-text and Speech for text-to-speech.
- Both models are described by Mistral as state of the art and serve as the engine of the company's new voice agents API, which comes with an SDK built on the open-source LiveKit framework.
- Voxtral 3 Speech supports more than 35 languages, and Pro Voice adds cloning that reproduces a voice from roughly 30 minutes of recordings.
- There is no dedicated Voxtral 3 page or standalone app yet, so access runs through the API; pricing, latency and cloning safeguards are not published, so confirm them on the official site, mistral.ai.
Quick Facts
| Type | Speech-to-text and text-to-speech models via API |
|---|---|
| Devices | Cloud API; any device with a network connection |
| Price | Not published; check mistral.ai for current rates |
| Languages | Voxtral 3 Speech: more than 35 languages |
| Highlight | Pro Voice clones a voice from about 30 minutes of audio |
| Learning curve | Moderate for developers; SDK handles audio transport |
| Category | AI & Models |
| Creator | Mistral AI |
| Official site | mistral.ai ↗ |
| Heat | 🔥 274.4K Heat |
| Last updated |
What Mistral AI's Voxtral 3 Actually Is
Voxtral 3 is the newest generation of voice models from Mistral AI, the company behind the Mistral family of large language models. Announced on October 20, 2026, it arrives not as one model but as a pair: Voxtral 3 Transcribe, which turns spoken audio into text, and Voxtral 3 Speech, which generates spoken audio from text. Mistral describes both as state of the art. Neither is pitched as a standalone consumer app; together they form the engine of the company's new voice agents API.
That framing matters. Most people searching for Voxtral 3 are not looking for a dictation app or a novelty voice generator. They are developers, product teams and founders who want assistants that listen to a person and answer out loud, without stitching together three vendors for transcription, synthesis and real-time audio transport. Mistral's pitch is that the whole loop now comes from one place, with an SDK built on LiveKit handling the audio plumbing.
There is no dedicated product page for Voxtral 3 yet, so the practical entry point is the official site, mistral.ai, where API access and documentation live. If you want to evaluate it today, plan around the API rather than a downloadable app. Details such as exact rate limits, regional availability and per-minute pricing have not been published in the announcement, so confirm them on the official site before you commit a production workload.
- Two models, one release
Transcribe handles speech-to-text while Speech handles text-to-speech. Mistral presents them as a matched pair rather than separate products.
- Built for agents, not apps
The models are the engine of Mistral's voice agents API, aimed at assistants that hold spoken conversations with people.
- Announced October 20, 2026
Mistral AI introduced the generation as state of the art for both transcription and speech synthesis.
- Check the official site first
No dedicated Voxtral 3 page exists yet, so mistral.ai is where API access and current limits are documented.
Inside the Pair: Transcribe, Speech and Pro Voice Cloning
Voxtral 3 Transcribe is the listening half. Feed it recorded or streaming audio and it returns text, which is the foundation for captions, meeting notes, call analytics and the input stage of any voice agent. Mistral calls it state of the art, though the announcement does not publish benchmark tables, so treat comparative accuracy claims as vendor positioning until independent evaluations appear. What is clear is the intended role: reliable transcription that feeds directly into the same stack as the speaking model.
Voxtral 3 Speech is the speaking half, and it supports more than 35 languages. That breadth is the headline for anyone building for a global audience, because a single agent can greet a caller in one language and switch to another without a separate provider. Pro Voice sits on top of Speech and adds voice cloning: supply roughly 30 minutes of recordings and the model reproduces that voice. For brands, that means one consistent, recognizable sound across every interaction.
Voice cloning deserves a careful read. Thirty minutes of audio is a modest amount to collect, which makes Pro Voice practical for a company recording its own spokesperson, and equally raises consent and impersonation questions. Mistral has not detailed safeguards, watermarking or verification requirements in the announcement. If you plan to clone a voice, get explicit written permission from the speaker, keep the recordings secure, and check the official site for the current acceptable-use terms before you build anything customer-facing.
- Transcribe: audio in, text out
It covers the listening side of the loop, from recorded files to live streams feeding an agent.
- Speech: text in, audio out
It generates spoken output in more than 35 languages from a single model, which suits multilingual products.
- Pro Voice cloning
Roughly 30 minutes of recordings is enough to reproduce a voice for a consistent agent identity.
- Unconfirmed details
Benchmarks, latency figures and safety tooling are not in the announcement, so verify them on the official site.
Building a Voice Agent With the API and the LiveKit SDK
The voice agents API is where the two models become a product. It ships with an SDK built on LiveKit, the open-source framework for real-time audio and video. In practice that means you can connect a microphone, a phone call or a browser audio stream to an agent without writing your own transport layer, buffering logic and interruption handling. LiveKit is widely used for real-time media, so teams already familiar with it will recognize the shape of the integration.
A typical build looks like this: audio arrives through LiveKit, Transcribe converts it to text, your language model decides what to say, and Speech renders the reply as audio that streams back to the caller. Because both ends come from Mistral, you avoid mismatched formats and separate billing relationships. The trade-off is concentration: if the API has an outage or a latency spike, both halves of your conversation are affected at once.
Start small. Prototype with recorded audio before you touch live calls, measure round-trip latency on your own network, and test how the agent handles interruptions, silence and background noise. Voice feels fast to users even when it is technically slow, so a few hundred milliseconds of delay can make an otherwise accurate agent feel unnatural. Tune your turn-taking logic early, and keep a fallback path in case the API is unavailable.
- LiveKit SDK
The SDK handles real-time audio transport, so you focus on agent logic instead of media plumbing.
- One vendor, one loop
Transcription and synthesis share a stack, which simplifies formats, credentials and billing.
- Latency is the real metric
Measure round-trip delay on your own network, because perceived speed matters more than raw accuracy.
- Plan for failure
Design a fallback for outages, since a single provider covers both listening and speaking.
How Voxtral 3 Compares, What It Cannot Do and Who It Suits
Comparable ground is crowded. Other major AI companies and several specialist speech providers all offer transcription and synthesis APIs, and many teams already combine a transcription service with a separate text-to-speech provider. Voxtral 3's argument is integration: one company for hearing, speaking and the real-time agent layer, with cloning included. Whether that beats a best-of-breed stack depends on your priorities. If you need a specific voice, a niche language or a certified compliance posture, check whether Voxtral 3 covers it before switching.
The limitations are mostly about what has not been published. There is no dedicated Voxtral 3 page yet, no public benchmark table, no confirmed pricing in the announcement and no detail on how Pro Voice cloning is policed. Latency, concurrency limits and data-retention terms are also unstated. None of that means the models are weak; it means you should treat marketing language as a starting point and validate with your own audio, in your own languages, on your own infrastructure.
Who is it for? Developers building voice agents, contact-center tools, accessibility features, language-learning products and multilingual customer support. It is less obviously aimed at hobbyists who want a quick text-to-speech toy, since there is no standalone app. If you are a solo builder, the API plus LiveKit SDK is still approachable, but budget time for audio testing. If you are an enterprise, the single-vendor loop is attractive, provided the compliance answers you need appear on the official site.
- Versus best-of-breed stacks
You trade some flexibility for one integration, one bill and a matched pair of models.
- Versus standalone speech apps
Voxtral 3 is API-first, so there is no consumer app to download and click through.
- Open questions
Pricing, latency, retention and cloning safeguards are unconfirmed, so verify them before production.
- Best fit
Teams shipping multilingual voice agents, especially those already comfortable working with LiveKit.
Step-by-Step Guide
Confirm access on the official site
Go to mistral.ai and look for the voice agents API and Voxtral 3 documentation. Because no dedicated product page exists yet, the official site is the only reliable source for current model names, availability, rate limits and pricing. Note what is documented and what is still missing.
Get API credentials and choose a model
Create or sign in to your Mistral account and generate an API key. Decide which half you need first: Transcribe if you are turning existing audio into text, Speech if you need generated voice, or both if you are building a conversational agent. Keep keys server-side, never in client code.
Test with recorded audio before going live
Send a handful of real recordings through Transcribe and compare the output against your own transcript. For Speech, generate the same sentence in several of the supported languages and listen for pronunciation, pacing and consistency. This baseline tells you whether the models suit your accents and domain vocabulary.
Wire the LiveKit SDK into your agent loop
Install the SDK and connect your audio source, then chain the pieces: audio in, transcription, your language model's response, speech synthesis, audio out. Handle interruptions and silence explicitly. Test on the networks and devices your users actually have, including mobile connections with variable quality.
Add Pro Voice cloning carefully
If you need a consistent brand voice, collect around 30 minutes of clean recordings from a speaker who has given written consent, then configure Pro Voice. Review the acceptable-use terms on the official site first. Keep the source recordings secured and document who authorized the clone and for which use cases.
Frequently Asked Questions
What is Voxtral 3?
Voxtral 3 is Mistral AI's newest generation of voice models, announced on October 20, 2026. It comes as two parts: Voxtral 3 Transcribe, which converts speech to text, and Voxtral 3 Speech, which generates speech from text in more than 35 languages. Together they power Mistral's voice agents API.
Is Voxtral 3 free, and how much does it cost?
Mistral has not published Voxtral 3 pricing in the announcement, and there is no dedicated product page yet. Access runs through the voice agents API rather than a free consumer app. Check the official site, mistral.ai, for current rates, free-tier allowances and any usage limits before you plan a budget.
Can I use Voxtral 3 without being a developer?
Realistically, no. Voxtral 3 is API-first, with an SDK built on LiveKit for real-time audio. There is no standalone app to download, so using it means writing or adapting code, or relying on a product that has already integrated the API. Non-developers should wait for third-party tools or a consumer release.
How does Pro Voice cloning work, and is it safe?
Pro Voice reproduces a voice from roughly 30 minutes of recordings, giving an agent one consistent sound. Mistral has not detailed safeguards, watermarking or verification in the announcement. Only clone voices with explicit written consent, store recordings securely, and review the acceptable-use terms on the official site before any customer-facing deployment.
How does Voxtral 3 compare with other speech models?
Its main differentiator is bundling: transcription, synthesis and a real-time agent SDK from one vendor, with cloning included. Competing options from other AI companies and specialist speech providers may win on a specific voice, language or compliance certification. Because Mistral has not published benchmarks for Voxtral 3, test both on your own audio before deciding.
