Mistral Voxtral 3: voice agent SDK and speech models

What did Mistral release?

Mistral introduced Voxtral 3, a new family of speech-to-text and text-to-speech models. It ships alongside a voice agent SDK that combines listening, reasoning, tool calling and speaking in one realtime streaming pipeline (Mistral). The targets are contact centers, service desks, virtual assistants and hands-free copilots.

What can transcription do?

Voxtral Small Transcribe 3 supports more than 50 languages, including all 24 official EU languages. It handles diarization (who spoke when) and context biasing, where supplied key terms improve recall on domain vocabulary without fine-tuning. It also detects when a speaker has finished, so agents can respond without delay or premature cut-offs.

On Mistral's own FLEURS evaluation across 45 languages, Small reaches roughly 4% word error rate, while Whisper large-v3 sits between 15% and 16%. Mistral says it also beats Scribe v2 and Gemini 3.5 Transcribe. In diarization it leads the chart with an 18.4% cpWER. The Mini variant has 4B parameters, and Realtime Transcribe 3 transcribes at under 500 ms latency in more than 35 languages.

How does it speak?

Voxtral Speech 3 covers more than 35 languages and 50 dialects. It infers emotion, pacing and emphasis from context, and these can also be set with instructions. Voice cloning is included, and Pro Voice reproduces a speaker's character and style across languages from 30 minutes of audio. Voice presets are available in 18 languages.

What does the SDK do?

The voice agent SDK is built on LiveKit. Realtime Transcribe 3 listens, a general model such as Mistral Small reasons and calls tools, and Speech 3 replies. No stage waits for the previous one, so according to Mistral each turn completes in under a second. The agent stops talking when the caller cuts in.

During a call the agent can verify identity, check a bill or book an appointment using function calling, web search and retrieval over your own systems. It runs over SIP for phone calls and WebRTC in browsers and apps. Each stage is a separate model, so the LLM, the voice or the language can be swapped without rewriting the agent.

Pricing and availability

Voxtral 3 is available now through the Mistral API on European, US and global endpoints, and it can be tested in the Studio playgrounds. Mini Transcribe 3 costs $0.0035 per minute in batch, Small Transcribe 3 $0.004 and Realtime Transcribe 3 $0.008 per minute. Speech 3 costs $13 per million characters. The full stack can also run on-premises or in a private cloud through Mistral's sales team.

For European teams, EU hosting and self-deployment make a stronger privacy case for customer-service voice agents. The benchmarks are Mistral's own, so test on your own audio before going to production.