Mistral launches the Voxtral 3 speech model family and a voice agent SDK: transcription in 50+ languages, TTS at $13 per 1M characters, EU-hosted or self-deployed

Background and pain points

A voice agent has to listen accurately across languages, accents and interruptions, understand intent, speak naturally, and take multi-step actions, all in real time. Teams usually stitch together ASR, an LLM and TTS from different vendors, which stacks latency and often sends audio across borders. Mistral now ships the models plus orchestration, hosted in the EU or self-deployed, aimed at contact centers, service desks and hands-free copilots.

Architecture highlights

Voxtral 3 is a family of separate models: Small Transcribe 3 (highest-accuracy STT), Mini Transcribe 3 (4B parameters, low cost), Mini Transcribe Realtime 3 (sub-500 ms streaming STT) and Mini TTS 3, also called Voxtral Speech 3. The STT models support diarization, context biasing and end-of-turn detection that uses both acoustic and semantic cues. TTS infers emotion, pace and emphasis from context and also accepts tags such as [emotion=sad][pace=slow]. The LiveKit-based SDK streams every stage without waiting on the previous one; Mistral says each turn completes in under a second and the agent stops when the caller cuts in.

Benchmarks

All figures are vendor-reported and not yet independently reproduced. FLEURS 45-language macro WER: Small 4.1%/4.0% (tied with Microsoft AI Transcribe 2), Mini 4.6%/4.5%, Gemini 3.5 Transcribe 6.1%/4.5%, Whisper large-v3 16.0%/14.9%. Diarization cpWER: Small 18.4% vs 30.9% for the previous Mini Transcribe 2. Earnings-22 keyphrase recall with context: Small 94.6% vs Scribe v2 94.3%. In realtime, the 480 ms setting scores 6.4% WER, behind Gemini 3.5 Realtime Transcribe at 4.2%. TTS charts plot quality per dollar judged by gemini-3.8-flash, mostly internal evals, without per-model scores published.

Getting started

Try voice agents, transcription, speech and voice cloning in Mistral Studio playgrounds; the API has EU, US and global endpoints, with transcription at /v1/audio/transcriptions. Mini Transcribe 3 costs $0.0035/min and Speech 3 $13 per 1M characters; see the pricing page for the rest. Phone agents can connect over SIP through the SDK, and teams needing on-prem can deploy the full stack with Mistral. Mini Transcribe 3 is currently marked Public Preview in the docs.