Mistral's Voxtral 3 pairs multilingual speech models with a real-time voice agent SDK

Mistral has released Voxtral 3, a family of speech-to-text and text-to-speech models accompanied by a voice agent SDK that folds listening, reasoning, tool calling and speaking into a single real-time streaming pipeline. According to Mistral, whose announcement reached the Hacker News front page, the release targets contact centers, service desks, virtual assistants and hands-free copilots, with transcription support for more than 50 languages.

One pipeline instead of stitched-together parts

Mistral frames the problem this way: transcription and voice generation have each improved on their own, but a reliable voice agent has to do four things at once — listen accurately across languages, accents and interruptions, reason about what a user means in context, speak naturally with expression, and take multi-step actions across connected systems. Getting all four to work together in real time is still hard.

The SDK is the company's answer. Agents built with it can retrieve information, interact with business systems and complete tasks during a live conversation over phone or web. End-of-turn detection is built into the transcription models so agents know when a user has finished speaking and can respond without delay.

What is in the family

Voxtral Small Transcribe 3 is the multilingual speech-to-text flagship, covering 50 languages including 24 official EU languages and a number of low-resource Asian and Indic languages. It supports diarization, end-of-turn detection and context biasing, where key terms or free-form text are supplied to the model to improve recall of domain-specific vocabulary without fine-tuning.

Voxtral Mini Transcribe 3 is a 4-billion-parameter variant priced at $0.0035 per minute, positioned for fast and inexpensive transcription. A realtime model, Voxtral Realtime Transcribe 3, is built for agentic use and delivers transcriptions in under 500 ms across more than 35 languages, with diarization and context biasing intact.

On the synthesis side, Voxtral Mini TTS 3 introduces voices that adjust emotion, pace and emphasis based on conversational context, supporting 35+ languages and 50+ dialects. A Pro Voice adaptation option reproduces a speaker's character, emotional range and speaking style across languages from 30 minutes of reference audio.

The numbers, as reported by Mistral

All benchmark figures come from Mistral itself, so they should be read as vendor claims. On FLEURS — macro-averaged over the 45 languages supported by all seven compared models — Voxtral Small Transcribe 3 posted a 4.1% word error rate with the language forced and 4.0% with it inferred, level with Microsoft AI Transcribe 2 (4.1%/4.0%) and ahead of Gemini 3.5 Transcribe (6.1%/4.5%), Scribe v2 (4.9%/4.9%), GPT Transcribe (6.3%/5.6%) and Whisper large-v3 (16.0%/14.9%).

On diarization, measured as concatenated minimum permutation WER over public benchmarks including AMI, Switchboard and AISHELL-4, Voxtral Small scored 18.4%, ahead of Pyannote Orchestrator (19.0%), Microsoft AI Transcribe 2 (20.3%) and Scribe v2 (23.0%). In keyphrase recall on Earnings 22, Voxtral Small reached 94.6% when context was supplied versus 81.4% without — a larger uplift from contextual prompting than Scribe v2 (94.3%/84.0%) or Gemini 3.5 Transcribe (87.3%/80.9%).

The realtime model shows the expected latency-accuracy tradeoff in Mistral's FLEURS 35 chart: roughly 5.7% WER at a 2,400 ms setting versus 7.4% at 240 ms, while Gemini 3.5 Realtime Transcribe sits at 4.2% and the non-streaming Voxtral Mini Transcribe 3 leads at 4.0%.

Deployment and sovereignty

As with Mistral's other models, Voxtral 3 is designed for sovereign deployment. It is available through Mistral's EU-hosted API or for self-deployment on a customer's own infrastructure, giving organizations control over where audio is processed, how data is handled and whether conversations are retained.

Why it matters

Voice is becoming the next interface battleground for AI agents, and Voxtral 3's bet is that developers want an integrated stack rather than a do-it-yourself assembly of separate transcription, LLM and TTS vendors. Tool calling inside the SDK, combined with sub-500 ms realtime transcription, is what turns speech models into agents that can act during a live call rather than just transcribe it. The sovereignty angle — EU hosting or full self-deployment — differentiates Mistral from US-centric rivals and matters for contact centers handling personal audio data. The crowded field in Mistral's own benchmarks, spanning Microsoft, Google, OpenAI, Whisper and Deepgram, shows how competitive speech AI has become; if the numbers hold up under independent testing, Voxtral 3 becomes a serious option for anyone building voice agents in or for Europe.