Mistral launches Voxtral 3: 50-language transcription, expressive speech in 35+ languages and a LiveKit-based voice agent SDK, API and self-hosted only

Why it matters

【Key points】

Mistral released Voxtral 3 on Oct 20 (US time): new speech-to-text and text-to-speech models plus a voice agent SDK, available the same day through the Mistral API (EU, US and global endpoints).

1. Transcription: Voxtral Small Transcribe 3 covers 50 languages (all 24 official EU languages plus low-resource Asian and Indic ones), with diarization, context biasing and end-of-turn detection. Mistral says it beats Whisper, Scribe v2 and Gemini 3.5 Transcribe on multilingual transcription and diarization. The Mini variant is 4B parameters at $0.0035/min batch; Small is $0.004/min.

2. Realtime: Voxtral Mini Transcribe Realtime 3 runs under 500 ms latency in 35+ languages, at $0.008/min.

3. Speech: Voxtral Mini TTS 3 (Voxtral Speech 3 in the API) covers 35+ languages and 50+ dialects, adapts emotion, pace and emphasis to context, and accepts tags like [emotion=sad]. It costs $13 per 1M characters. Pro Voice clones a voice from about 30 minutes of audio.

4. Voice agent SDK: built on LiveKit, it streams listening, reasoning and tool calls (by a general model such as Mistral Small) and speech as one pipeline. Mistral says each turn completes in under a second and the agent stops when the caller interrupts. SIP and WebRTC are supported.

5. Deployment: Mistral's EU-hosted API or self-deployment. The post does not mention open weights, and no Voxtral 3 weights are on Hugging Face.

【Verifiable】

Announcement: https://mistral.ai/news/voxtral-3

Launch post: https://x.com/MistralAI/status/2112574615318876178

【Our read】

1. The news is the full package: transcription, reasoning and speech bundled for phone agents, taking on OpenAI, ElevenLabs and Deepgram voice-agent offerings.

2. The state-of-the-art and price comparisons are Mistral's own evals, several judged by another lab's model; third-party results are still needed.

3. Unlike earlier open Voxtral releases, this generation so far ships only as an API and private deployment.