Mistral AI Launches Voxtral 3 Speech Models and Voice Agent SDK

Mistral AI has released Voxtral 3, a new family of speech-to-text and text-to-speech models paired with a voice agent SDK, the company announced.

According to Mistral, the release is aimed at building voice agents that can listen, reason, speak, and take multi-step actions through connected systems in real time. The company says existing speech tools have improved but that combining accurate listening, contextual reasoning, natural speech, and tool-calling into one reliable real-time system has remained difficult.

What's in the release

Mistral says Voxtral 3 includes two core models and an SDK:

• Voxtral Small Transcribe 3 — a speech-to-text model the company describes as state-of-the-art in multilingual transcription, supporting diarization (identifying who said what), context biasing (using supplied keyphrases to improve accuracy), and end-of-turn detection.
• Voxtral Mini TTS 3 — a text-to-speech model the company says produces voices that adjust emotion, pace, and emphasis based on conversational context, across 35+ languages and 50+ dialects.
• A voice agent SDK — which Mistral says combines listening, reasoning, tool calling, and speaking into a single real-time streaming pipeline, allowing agents to retrieve information and complete tasks during live phone or web conversations.

The company also describes a Pro Voice adaptation feature for professional use cases, which it says can reproduce a speaker's distinctive character, emotional range, and speaking style across languages using about 30 minutes of audio.

Language coverage and performance claims

Mistral says the transcription model now covers 50 languages, including 24 official EU languages and a number of low-resource Asian and Indic languages. The company states that the Small version achieves state-of-the-art results in diarization and multilingual transcription, including in English, French, and Spanish, and that it outperforms Whisper, Scribe v2, and Gemini 3.5 Transcribe on these measures. Mistral says a Mini variant, at 4 billion parameters, is priced at $0.0035 per minute.

The company also describes a separate model, Voxtral Realtime Transcribe 3, which it says delivers transcription at sub-500ms latency in more than 35 languages, including diarization and context biasing.

For speech generation, Mistral says Voxtral Mini TTS 3.0 is priced at $13 per 1 million characters, which the company states is less expensive than comparable offerings from OpenAI's TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash. Mistral says the model natively interprets the semantic meaning of text context to produce speech, and that new features include voice cloning from more than 30 minutes of sampled audio, custom prompt instruction following for emotion, pacing, and emphasis, and a vocabulary library.

Deployment and data handling

Mistral says Voxtral 3 is designed for sovereign deployment, consistent with the company's other models. It is available through Mistral's EU-hosted API or for self-deployment on an organization's own infrastructure, which the company says gives organizations control over where audio is processed, how data is handled, and whether conversations are retained.

Mistral AI, Voxtral 3: Voice Agent SDK & Speech Models in 50+ Languages, published October 20, 2026.