Three things in one release
Mistral AI released Voxtral 3 on October 20: speech-to-text models, a text-to-speech model and a voice agent SDK. Mistral's argument is that a usable voice agent has to do four things well at once: hear accurately across languages, accents and interruptions; understand what users mean in context; speak naturally with emotion; and carry out multi-step actions across connected systems. The new SDK chains listening, reasoning, tool calling and speaking into a single real-time streaming pipeline, aimed at contact centers, service desks, virtual assistants and hands-free copilots.
Transcription and speech specs
- Transcription: Voxtral Small Transcribe 3 supports 50 languages, including the 24 official EU languages and many low-resource Asian and Indic languages, with speaker diarization, context biasing (supply a list of terms to improve recall on proper nouns without fine-tuning) and end-of-turn detection. Mistral says it beats Whisper, Scribe v2 and Gemini 3.5 Transcribe on average word error rate across 45 FLEURS languages. The Mini version has 4 billion parameters and costs $0.0035 per minute. The realtime model, Voxtral Realtime Transcribe 3, runs at under 500 ms latency in more than 35 languages.
- Speech: Voxtral Mini TTS 3.0 covers more than 35 languages and 50-plus dialects, adjusts emotion, pace and emphasis to context, and costs $13 per million characters, which Mistral says is less than comparable offerings from OpenAI TTS-1, Deepgram, ElevenLabs Flash and Gemini Flash. Pro Voice can reproduce a speaker's character, emotional range and speaking style across languages from about 30 minutes of audio.
Deployment, and what still needs checking
Voxtral 3 is available through Mistral's EU-hosted API or for deployment on a company's own infrastructure, letting the company decide where audio is processed and whether it's retained. That's Mistral's familiar "sovereign deployment" pitch, and it matters in practice for European customer service and healthcare settings bound by data residency rules.
Note that every comparison in the announcement comes from Mistral's own testing, and some speech evaluations use gemini-3.8-flash as the judge, so independent reproduction is still to come. The better voice cloning gets, the greater the risk of impersonation; before adopting Pro Voice, companies should at least confirm consent from the person being cloned and settle a retention policy.