Mistral releases Voxtral 3: 50-language transcription, expressive TTS and a real-time voice agent SDK

Mistral AI has released Voxtral 3, a family of speech-to-text and text-to-speech models plus a voice agent SDK that runs listening, reasoning, tool calling and speaking in one real-time streaming pipeline. Transcription supports 50 languages, and the 4B-parameter Voxtral Mini Transcribe 3 is priced at $0.0035 per minute. The models are available now through Mistral's EU-hosted API or for self-deployment.

What's in the release

• Voxtral Small Transcribe 3: the larger speech-to-text model. It supports diarization (who spoke when), context biasing and end-of-turn detection. Parameter count and pricing are not yet disclosed.
• Voxtral Mini Transcribe 3: a 4B-parameter variant at $0.0035/min. It also supports context biasing.
• Voxtral Mini TTS 3: text-to-speech with voices that adjust emotion, pace and emphasis to conversational context. It covers 35+ languages and 50+ dialects. Pricing is not yet disclosed.
• Voice agent SDK: combines transcription, reasoning, tool calling and speech for live phone or web conversations. Mistral's demo shows an agent reading a French referral letter and booking a dermatology appointment through tool calls.
• Pro Voice: voice adaptation that Mistral says reproduces a speaker's character, emotional range and style across languages from 30 minutes of audio.

Language coverage grows to 50, including all 24 official EU languages and several low-resource Asian and Indic languages.

Benchmarks (published by Mistral)

Mistral claims Voxtral Small Transcribe 3 is state-of-the-art in multilingual transcription and diarization, and that it outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe. These figures come from Mistral's own charts and have not been independently verified here.

• Diarization (cpWER, lower is better, macro-average over AMI-IHM, AMI-SDM, Switchboard, Switchboard CallHome and AISHELL-4): Voxtral Small Transcribe 3 scores 18.4%, Voxtral Mini Transcribe 3 18.6%, Pyannote Orchestrator 19.0%, Microsoft AI Transcribe 2 20.3%, Scribe v2 23.0%, Voxtral Mini Transcribe 2 30.9% and Gemini 3.5 Transcribe 38.4%.
• Multilingual transcription (FLEURS, 45 shared languages, macro-average WER, lower is better): Voxtral Small Transcribe 3 lands at roughly 4.0–4.1%, depending on whether the language is specified or auto-detected. Whisper large-v3 is at about 15–16%.
• Keyphrase recall (Earnings22, higher is better): with contextual prompting, Voxtral Small Transcribe 3 reaches 94.6% (81.4% without context), against 94.3% for Scribe v2 (84.0% without) and 87.3% for Gemini 3.5 Transcribe (80.9% without).

The release also references an Artificial Analysis word-error-rate comparison, but the ranking data in the source is truncated, so no placement is reported here.

Deployment and data control

Voxtral 3 is offered through Mistral's EU-hosted API or for self-deployment on customer infrastructure. Mistral says this lets organizations control where audio is processed and whether conversations are retained. The source does not state weight licensing terms, context limits or the release date.

What this means

Mistral is moving from standalone speech models to a full voice-agent stack, which puts it in direct competition with ElevenLabs, OpenAI and Google. The EU-hosted and self-deploy options are the clearest differentiator for regulated sectors such as healthcare, banking and public services, where audio residency matters.

The benchmark claims need careful reading. On FLEURS, Microsoft AI Transcribe 2 appears to sit level with Voxtral Small Transcribe 3 in Mistral's own chart, though it is absent from the company's list of models it says it beats. The clearer wins are on diarization and Whisper-relative accuracy. The $0.0035/min Mini price is aggressive, but the missing pricing for Small and TTS makes full cost comparisons impossible for now. Independent evaluations of latency under interruption and tool-calling reliability will matter more than WER for production agents.