Mistral launches Voxtral 3 with voice agent SDK and new speech models

A single pipeline for voice agents

Mistral says that building voice agents that can listen, reason, interact and handle complex workflows remains difficult. In its view, a reliable agent must do four things at once: listen accurately across languages, accents, speakers and interruptions, reason about what users mean, speak naturally with emotion, and take action over several steps through connected systems.

Voxtral 3 is the company's answer. The new voice agent SDK combines listening, reasoning, tool calling and speaking in a real-time experience. According to Mistral, agents can retrieve information, interact with business systems and complete tasks during live conversations over phone or web. The target uses named are contact-center agents, service desks, virtual assistants and hands-free copilots.

Mistral illustrates this with a demo in which a caller asks an agent to book a French referral letter. The agent looks up the reference number with a tool call, translates the letter's request for a dermatological opinion, and books an appointment.

Transcription: Small, Mini and Realtime models

The speech-to-text side is led by Voxtral Small Transcribe 3. Mistral calls it state-of-the-art in multilingual transcription and says it supports diarization (identifying who said what), context biasing and end-of-turn detection. The supported language list has grown to 50, including 24 official EU languages and many low-resource Asian and Indic languages.

According to Mistral, the Small version achieves state-of-the-art results across diarization and multilingual transcription, including in English, French and Spanish, and outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe. The Mini variant has 4B parameters and costs $0.0035/min, the company says.

Both models support context biasing: key terms and free-form text can be supplied to improve recall on technical vocabulary. Mistral says this allows production use with domain-specific terms without fine-tuning.

Voxtral Realtime Transcribe 3 is aimed at agentic use. Mistral says it delivers high-quality transcription at sub-500ms latency in over 35 languages, including with diarization and context biasing. End-of-turn prediction uses both acoustic and semantic information to judge when a user has finished speaking, which Mistral says keeps latency low while limiting false cut-offs when users are interrupted prematurely.

Text-to-speech and voice adaptation

Voxtral Mini TTS 3 is available in batch and realtime modes. Mistral says it reaches state-of-the-art expressiveness and instruction following for 35+ languages and 50+ dialects, with voices that adjust emotion, pace and emphasis to the conversational context. Explicit instruction variables include emotion, pacing and emphasis.

On price, Mistral states that Voxtral Speech 3 costs $13 per 1M characters, which it says is less than comparable models from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash.

For professional cases where basic voice cloning is not enough, Mistral offers Pro Voice adaptation. The company says it can reproduce a speaker's character, emotional range and speaking style across languages from just 30 minutes of audio. Other listed features are custom prompt instruction following, contextual TTS and a vocabulary library.

What it means for organizations

Mistral stresses sovereign deployment. Organizations can use the EU-hosted API or run Voxtral 3 on their own infrastructure, which, according to the company, gives them control over where audio is processed, how data is handled and whether conversations are retained.

For businesses building phone or web voice assistants, the release bundles transcription, speech synthesis and agent tooling from one vendor, with published per-minute and per-character prices for the Mini models. The benchmark claims come from Mistral itself, so teams will want to test the models on their own languages and vocabularies.