Mistral releases Voxtral 3 speech models and a voice agent SDK

Mistral AI has announced Voxtral 3, a new family of speech-to-text and text-to-speech models, together with a voice agent SDK that joins listening, reasoning, tool calling and speaking in one real-time streaming pipeline. Mistral frames the problem this way: a reliable voice agent has to listen accurately across languages, accents, speakers and interruptions, reason about what users mean, speak with emotion and expression, and take multi-step action through connected systems, and getting all four to work together in real time is still hard. The target uses it names are contact-center agents, service desks, virtual assistants and hands-free copilots.

On transcription, the model is called Voxtral Small Transcribe 3 (the page also uses the name Voxtral Transcribe 3). The supported language list grows to 50, including 24 official EU languages and many low-resource Asian and Indic languages. Mistral says the Small version reaches state-of-the-art results in diarization (who spoke when) and multilingual transcription, including English, French and Spanish, and that it outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe; no benchmark figures appear in the available text. The Mini variant has 4B parameters and is priced at $0.0035 per minute. Both models support context biasing: key terms and free-form text can be supplied to improve recall on technical vocabulary, which Mistral says lets them handle domain-specific terms in production without fine-tuning. A separate model, Voxtral Realtime Transcribe 3, is built for agents and delivers transcription at sub-500ms latency in over 35 languages, with diarization and context biasing. End-of-turn prediction is built in and uses both acoustic and semantic information, which Mistral says allows low latency while limiting false cut-offs when a user is interrupted prematurely.

On speech output, Voxtral Mini TTS 3 (also written Mini TTS 3.0, Voxtral Speech 3 and Voxtral TTS) is available in batch and realtime and supports 35+ languages and 50+ dialects. Its voices adjust emotion, pace and emphasis from the conversational context, and the model reads the text context to pick emotion and voice style without extra instructions or voice prompts. New features include voice cloning from 30+ minutes of sampled audio, custom prompt instruction following, contextual TTS and a vocabulary library; explicit controls cover emotion, pacing and emphasis. Priced at $13 per 1M characters, Mistral says Voxtral Speech 3 costs less to run than comparable models from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. Voice Presets exist in 18 languages: Arabic (Emirati, Modern Standard), Brazilian Portuguese, Catalan, Mandarin Chinese, Dutch, English (American, British), French (France), German, Greek, Hindi, Italian, Korean, Norwegian, Russian, Spanish (Spain) and Vietnamese. A separate Pro Voice adaptation, aimed at professional cases where basic cloning is not enough, reproduces a speaker's character, emotional range and speaking style across languages from just 30 minutes of audio.

The page shows demos. One is a booking scenario: a caller named Béatrice asks the agent to book an appointment from a French referral letter; the agent reads the letter, finds a Tuesday slot at 5:45 pm with a dermatologist, calls a book_appointment tool, and confirms in English. Others contrast Voxtral 1 and Voxtral 3 output on hard cases: an email address, flight times and dates, a name, an angry complaint about a passport, and invented names such as Quivello and Xyra with phoneme hints.

The voice agent SDK is built on LiveKit. In the pipeline, Voxtral Mini Transcribe Realtime 3 listens and detects when the caller has finished speaking, a general model such as Mistral Small reasons and calls tools, and Voxtral Speech 3 replies in the tone of the conversation so far. Mistral says no stage waits for the one before it, so each turn completes in under one second, and the agent stops talking when the caller cuts in. During a call the agent can verify identity, check a bill or book an appointment through function calling, web search and retrieval over a company's own systems. It answers and places calls over SIP, runs in browsers and apps over WebRTC, and connects to existing telephony, IVR and contact-center software. Each stage is a separate model, so the LLM, voice or language can be swapped without rewriting the agent.

As with other Mistral models, Voxtral 3 is designed for sovereign deployment: organizations can use Mistral's EU-hosted API or deploy on their own infrastructure, and control where audio is processed, how data is handled and whether conversations are retained. It is available today through the Mistral API on European, US and global endpoints, and the Studio playgrounds let users try voice agents, transcription, speech generation and voice cloning.

Key facts

Voxtral 3 is a family of speech-to-text and text-to-speech models plus a voice agent SDK built on LiveKit, available today through the Mistral API on European, US and global endpoints.
Transcription covers 50 languages including 24 official EU languages; the Mini variant has 4B parameters and costs $0.0035 per minute, and the realtime model reaches sub-500ms latency in over 35 languages.
Voxtral Speech 3 (TTS) supports 35+ languages and 50+ dialects at $13 per 1M characters, which Mistral says is cheaper to run than OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash.
Mistral says the agent pipeline of listening, reasoning and speaking completes each turn in under one second, over SIP telephony or WebRTC, and each stage can be swapped.
Pro Voice adaptation reproduces a speaker's voice and style from 30 minutes of audio; the models can be used via the EU-hosted API or self-deployed.

Why it matters

Voice agents need transcription, reasoning, speech and tool use to work together in real time, and Mistral is offering all of those pieces from one vendor, with an SDK that ties them into a single streaming pipeline. The pitch is also about where the data goes: an EU-hosted API or self-deployment, with control over audio processing and retention. Mistral claims the Small transcription model beats Whisper, Scribe v2 and Gemini 3.5 Transcribe and that its TTS is cheaper to run than several named rivals.

Who it affects

Mistral names contact-center agents, service desks, virtual assistants and hands-free copilots as the intended uses. Organizations that care about where audio is processed and whether conversations are retained are the audience for the EU-hosted and self-deployed options. Teams working in European and low-resource Asian and Indic languages are called out in the language coverage.

How to use it

Voxtral 3 is available today through the Mistral API on European, US and global endpoints. The Studio playgrounds let you test voice agents, audio transcription, speech generation and voice cloning. Published prices: $0.0035 per minute for the Mini transcription model and $13 per 1M characters for Voxtral Speech 3. The voice agent SDK is built on LiveKit, handles calls over SIP and browser or app sessions over WebRTC, and connects to existing telephony, IVR and contact-center software. Because each stage is a separate model, the LLM, voice or language can be changed without rewriting the agent. Self-deployment is offered, but its terms are not covered in the available text of the announcement.

How solid is it

This is Mistral's own launch announcement, so every performance and cost claim is the company's. The comparison with Whisper, Scribe v2 and Gemini 3.5 Transcribe comes with no benchmark figures in the available text, and the same goes for the TTS cost comparison beyond the $13 price. No independent reviewers or customers are named. The page is also inconsistent in naming (Voxtral Transcribe 3 and Voxtral Small Transcribe 3; Voxtral Mini TTS 3, Mini TTS 3.0, Voxtral Speech 3 and Voxtral TTS) and in the audio needed for voice cloning (30 minutes versus 30+ minutes), and it does not explain these.

Risks and caveats

Prices are given only for the Mini transcription model and the TTS model; none is stated for the Small or realtime transcription models or for Pro Voice. The under-one-second turn time is Mistral's statement, with no conditions given. Licence and self-deployment details are not covered in the available text of the announcement. The demo comparisons between Voxtral 1 and Voxtral 3 are selected by Mistral.

“Each stage is a separate model, so you can change the LLM, the voice, or the language without rewriting the agent”

— Mistral AI, Voxtral 3 announcement