Voxtral 3: Mistral's Full-Stack Platform for Voice Agents

Building a voice agent means getting four things right at once: listening accurately, reasoning about intent, speaking naturally, and taking action across connected systems. Mistral's Voxtral 3 is an attempt to deliver all four in a single stack, with new transcription and speech models and an SDK that ties them together.

Listening: transcription in 50+ languages

Voxtral Small Transcribe 3 is Mistral's highest-accuracy transcription model. It covers 50+ languages, including the 24 official EU languages and many low-resource Asian and Indic ones. It also handles diarization (who said what) and context biasing, where you supply key terms or free-form text to improve recall on domain vocabulary without fine-tuning. Mistral reports it outperforming Whisper, Scribe v2, and Gemini 3.5 Transcribe.

For live use there is Voxtral Mini Transcribe Realtime 3, with sub-500ms latency in 35+ languages. It predicts end-of-turn from both acoustic and semantic cues, so agents respond quickly without cutting people off.

Speaking: expressive, local-sounding speech

Voxtral Speech 3 covers 35+ languages and 50+ dialects. It reads the context of the text to pick emotion, pacing, and emphasis, and you can also set these explicitly. Phoneme hints help with unusual names and words. The release also targets enterprise failure points like numbers and unique names, such as email addresses, prices, and dates. For brand voices, Pro Voice adapts a speaker's character and style across languages from about 30 minutes of audio.

Acting: the voice agent SDK

The SDK is built on LiveKit and runs speech in, reasoning, and speech out as one streaming pipeline. Each turn completes in under a second, and the agent stops talking when the caller interrupts. During a call it can use function calling, web search, and retrieval over your own systems, for example to verify an identity or book an appointment. It answers and places calls over SIP, runs in browsers and apps over WebRTC, and connects to existing telephony and contact-center software.

flowchart LR
  A[Caller audio] --> B[Realtime Transcribe 3]
  B --> C[Reasoning model + tools]
  C --> D[Voxtral Speech 3]
  D --> E[Spoken reply]

Each stage is a separate model, so you can swap the LLM, the voice, or the language without rewriting the agent.

Deployment and pricing

Voxtral 3 is available through Mistral's API on European, US, and global endpoints. To run the full stack on-premises or in a private cloud, you contact Mistral's team. This gives organizations control over where audio is processed and whether conversations are retained, which speaks directly to data-sovereignty requirements.

Worth keeping in mind

The benchmark charts are presented by Mistral itself. Several of the TTS comparisons are internal evaluations, and all of them are scored by an LLM judge rather than human listeners. Results also vary by chart: on the FLEURS chart, for example, Small Transcribe 3 is roughly level with Microsoft AI Transcribe 2. Treat the numbers as a strong starting point and test on your own audio, accents, and vocabulary before committing.

Even so, the combination of low latency, multilingual coverage, and sovereign deployment makes Voxtral 3 a serious option for contact centers, virtual assistants, and hands-free copilots.

Source: Mistral's Voxtral 3 announcement