Voxtral 3

Voxtral 3 introduces a cutting-edge sound delegate SDK and state-of-the-art address models designed for real-time, multi-step conversations throughout 50+ languages. It combines exact transcription, natural address synthesis alongside affectional adaptability, and profound tool integration for seamless workflow automation. Available via Mistral’s EU-hosted API or self-deployment, it ensures data sovereignty during delivering low-latency, human-like interactions for communication centers, virtual assistants, and hands-free copilots.

Speech transcription and sound generation have arrive a lengthy way, but construction end-to-end sound agents that can listen, reason, interact, and grip complex workflows remains difficult. 

A dependable sound delegate needs to do four things extremely fine (and simultaneously): 

1. Listen accurately across languages, accents, speakers, and interruptions,

2. Reason deeply about what users average in particular contexts,

3. Speak naturally with emotion and expression, and 

4. Take act over multiple steps through connected systems. 

Getting all four to activity together reliably in genuine period remains a challenge. 

To location this, we current Voxtral 3: a new family of state-of-the-art speech-to-text and text-to-speech models, alongside a sound delegate SDK that brings listening, thinking, tool calling, and speaking into a sole real-time streaming pipeline.

Voxtral 3 is built to create natural conversational experiences, from contact-center agents and assistance desks to virtual assistants and hands-free copilots. With assistance for dozens of languages, expressive speech, complex reasoning, and profound integration alongside your own tools and data, agents built alongside Voxtral can construe nuanced intent, logic through advanced requests, and execute multi-step actions throughout connected systems.

As alongside another Mistral models, the Voxtral 3 family is designed for sovereign deployment. Available today through Mistral's EU-hosted API or for self-deployment on your own infrastructure, it gives organizations authority complete anywhere audio is processed, how data is handled, and whether conversations are retained.

Highlights

1. Accurate transcription throughout 50+ languages. The speech-to-text model, Voxtral Small Transcribe 3, is state-of-the-art in multilingual transcription, identifying who stated what (diarization), using keyphrases to enhance its accuracy (context-biasing). It detects whenever person has completed speaking so agents can react without postpone (end-of-turn detection).

2. Natural address that adapts to the conversation. The text-to-speech model, Voxtral Mini TTS 3, introduces new voices that modify their emotion, pace, and accent according to conversational context, alongside assistance for 35+ languages and 50+ dialects.

3. Human-like intellect at your (spoken) command. The new sound delegate SDK combines listening, reasoning, tool calling, and speaking into a real-time experience. Agents can recover information, engage alongside endeavor systems, and complete tasks during live conversations complete phone or web.

4. Pro Voice for distinctly exact sound adaptation. For expert use cases anywhere essential sound cloning does not suffice, Voxtral’s Pro Voice adaptation is capable to reproduce a speaker’s distinctive character, affectional range, and speaking manner throughout languages, using fair 30 minutes of audio. 

5. Your sound AI, on your infrastructure. Use Mistral's EU-hosted API or deploy Voxtral 3 on your own infrastructure. Keep delicate conversations inside your surroundings and authority whether any audio is retained. 

Try Voxtral now  | Learn more

It hears 50+ languages, alongside state-of-the-art accuracy.

Voxtral Transcribe 3 represents a important milestone throughout transcription, diarization (who said when) and environment biasing capabilities. The supported catalog of languages is expanded to 50, including 24 authoritative EU languages and many low-resource Asian and Indic languages.

The Small type achieves state-of-the-art results throughout diarization and multilingual transcription, including in English, French and Spanish. It outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe. The Mini variant is 4B parameters, enabling accelerated and exact transcriptions at a cost of $0.0035/min.

Both models assistance environment biasing, anywhere key conditions and free-form content can be stated to bias the transcription and enhance recall on specialized terms. This enables them to be used in manufacturing settings alongside domain-specific vocabulary without the need for fine-tuning.

Voxtral Realtime Transcribe 3 is a realtime transcription example built for agentic use-cases. It achieves high-quality transcriptions at sub-500ms latency in complete 35 languages, including alongside diarization and environment biasing.

End-of-turn prediction is natively supported, incorporating the two acoustic and semantic data to accurately foretell whenever a person is completed speaking. This allows for low-latency sound delegate applications, during minimizing false cut-off rates whenever the person is prematurely interrupted.

It answers alongside the correct emotion, in your or your customer’s brand's own sound and local dialect.

Voxtral Mini TTS 3.0, accessible in lot and realtime, achieves state-of-the-art expressiveness and education following for 35+ languages and 50+ dialects. At $13 per 1M characters, Voxtral Speech 3 expenses small to run than comparable models from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash.

The address example implicitly interprets the semantic definition of environment better, producing natural address throughout a varied range of emotions and languages. New and improved features contain sound cloning from 30+ minutes of sampled audio, tradition immediate education following (IF), contextual TTS, and a vocabulary library, all of which assist the goal of address reliability in deployment. Variables for Explicit IF contain emotion, pacing, and emphasis.

All natural-sounding address relies on emotion, which determines its prosody, rhythm, pacing, emphasis, and transitions. Voxtral TTS natively interprets a prompt’s content environment to decide the emotion and sound style, making the example inherently improved at natural address without requiring additional instructions or sound prompts.

Illustrating Voxtral hill-climbing in expressiveness, numbers and distinctive names.

Speech appears local in 35 languages and complete 50 dialects.

Voxtral TTS supports complete 35 languages and local dialects. Our multilingual evals adhere to dialect differences in pronunciation, ceremonial vs. informal register, spelling, verbosity, and locale conventions (time, dates, temperatures, addresses, currencies).

Voxtral 3 multilingual library

Voice Presets are accessible in 18 languages: Arabic (Emirati, Modern Standard), Brazilian (Portuguese), Catalan, Chinese (Mandarin), Dutch, English (American, British), French (France), German, Greek, Hindi, Italian, Korean, Norwegian, Russian, Spanish (Spain), and Vietnamese.

The sound delegate is in authority of the entire conversation.

Voxtral 3 introduces a sound delegate SDK, built on LiveKit, that runs address in, reasoning, and address out as one streaming pipeline. Voxtral Mini Transcribe Realtime 3 listens and detects whenever the visitant has completed speaking, a broad model—like Mistral Small—reasons and calls tools, and Voxtral Speech 3 replies in the spirit of the conversation so far. No phase waits for the one before it, so all rotate completes in under one second, and the delegate stops talking whenever the visitant cuts in.

During a call, the delegate can verify a caller's identity, inspect a bill, or publish an appointment, using function calling, web search, and retrieval complete your own systems. It answers and places calls complete SIP, runs in browsers and apps complete WebRTC, and connects to the telephony, IVR, and contact-center application you already use.

Each phase is a distinct model, so you can alter the LLM, the voice, or the tongue without rewriting the agent, and run the entire stack on Mistral's API in Europe or on your own infrastructure.

Get started.

Voxtral 3 is accessible today through the Mistral API, hosted through our European, US, and earth endpoints.

You can test out sound agents, audio transcription, address generation, and sound cloning in the Studio playgrounds. 

To run the complete stack on-premises or in a personal cloud, talk to our team.