Voxtral 3: Mistral ships a voice agent SDK with 50-language transcription and emotional TTS

Mistral AI released Voxtral 3 on October 20, 2026. The release adds new speech-to-text and text-to-speech models plus a voice agent SDK that handles listening, reasoning, tool calls and spoken replies in one real-time stream. You can use it today through Mistral's EU-hosted API or deploy it on your own infrastructure.

What's new in Voxtral 3

Voxtral 3 is a family of models, not a single one. Here are the parts Mistral describes in its announcement:

• Voxtral Small Transcribe 3: the main speech-to-text model. It covers 50 languages, including all 24 official EU languages and a number of lower-resource Asian and Indic languages. It labels which speaker said what (diarization), accepts key terms to improve accuracy (context biasing), and detects when a speaker has stopped so an agent can answer without a long pause (end-of-turn detection).
• Voxtral Mini Transcribe 3: a smaller 4-billion-parameter transcription model built for speed and lower cost.
• Voxtral Realtime Transcribe 3: a streaming model for agent use. Mistral says it transcribes in more than 35 languages with latency under 500 ms, and it keeps diarization and context biasing.
• Voxtral Mini TTS 3: a text-to-speech model that comes in batch and real-time versions. It supports more than 35 languages and 50+ dialects, and its voices shift emotion, pacing and emphasis to fit the conversation. Developers can also set emotion, pacing and emphasis directly as instructions.
• Pro Voice: a voice adaptation option for professional use. Mistral says it can copy a speaker's character, emotional range and style across languages from about 30 minutes of audio.
• Voice agent SDK: ties all of this together so an agent can look up information, connect to business systems and finish tasks during a live phone or web call.

In the announcement's demo, an agent reads a French referral letter through a tool call, explains it to an English-speaking caller, and books an appointment.

How Voxtral 3 compares, according to Mistral

The benchmarks below come from Mistral itself. Mistral says the Small transcription model beats Whisper, Scribe v2 and Gemini 3.5 Transcribe on multilingual transcription and diarization. On its FLEURS chart covering 45 languages, Voxtral Small Transcribe 3 posts the lowest word error rate of the seven models compared. Vietnamese is one of the 45 languages in that comparison.

On speech output, Mistral says Voxtral Mini TTS 3 is cheaper to run than similar offerings from OpenAI, Deepgram, ElevenLabs and Google's Gemini Flash. It backs this with quality-per-dollar charts, some built on internal evaluations. For exact rates, see Mistral's pricing page.

The headline change from earlier Voxtral releases is the SDK. Before, developers had to wire transcription, a language model, tools and speech synthesis together on their own. Voxtral 3 puts those steps into one streaming pipeline.

Who should care and what to do

• Teams building contact-center bots, service desks or voice assistants: Mistral aims Voxtral 3 directly at these uses. The SDK and end-of-turn detection target the slow, stop-start feel that many voice bots have.
• Organizations with data residency rules: you can choose between the EU-hosted API and self-deployment, and you decide whether audio is stored. That matters if calls contain sensitive data.
• Developers transcribing specialist vocabulary: context biasing lets you pass in product names, medical terms or speaker names without fine-tuning.
• Anyone who only wants a local, open-weight model for personal use: check the license and deployment terms before you commit. Mistral describes self-deployment as an option for organizations, and the announcement does not say whether the weights are publicly downloadable.

How to try Voxtral 3

The announcement points to two places to start: the API keys section of the Mistral console (console.mistral.ai), where you create a key, and the audio section of Mistral's documentation (docs.mistral.ai), which has model pages for each transcription and TTS variant.

What Hacker News is saying

On Hacker News, the biggest complaint is that the models appear to be closed. Several commenters said that makes them less interested, especially for personal use where they don't want voice data leaving their machine. Some hoped for smaller models that run locally. Others questioned whether the zero-lag conversation in the demo reflects real use. One commenter said built-in context biasing is useful for rolling out custom solutions without self-hosting, and another praised an earlier small Voxtral model.

Frequently asked questions

How many languages does Voxtral 3 support?

The transcription models cover 50 languages. Real-time transcription covers more than 35, and the TTS model supports 35+ languages and 50+ dialects.

Can I run Voxtral 3 on my own servers?

Yes. Mistral offers Voxtral 3 through its EU-hosted API or as a self-deployment on your own infrastructure, and you control whether audio is retained.

What does the voice agent SDK do?

It combines listening, reasoning, tool calling and speaking in one real-time pipeline, so agents can take actions in connected systems during a live call.

How much audio does Pro Voice need?

Mistral says Pro Voice can reproduce a speaker's style across languages from about 30 minutes of audio.