Mistral Releases Voxtral 3: Voice Stack for Real-Time Agents

French AI company Mistral AI has announced the release of Voxtral 3, a series of speech recognition and synthesis models alongside a toolkit to develop voice assistants powered by LiveKit, a real-time audio/video framework. The speech recognition model supports over 50 languages with less than 500ms latency. It can be deployed either on your own servers or via Mistral’s EU-based cloud service. Recognising a minute of audio will cost you from $0.0035.

Low latency and the ability to deploy the solution on-premise are key factors when deciding if you can trust a voice agent to handle a live call with your customers. A half-second delay is enough to make the interaction feel natural rather than robotic, with awkward silences. According to Mistral, their speech recognition model achieves an error rate of 4–4.1% on the FLEURS 45 multilingual speech recognition benchmark. Low cost per minute makes it easier to get started with your first voice assistant for customer support or booking. And being able to deploy the solution on-premise gives you freedom from relying on a specific cloud provider. Is Voxtral 3 better than Whisper (OpenAI’s speech recognition model), or other solutions currently used in practice? That remains to be seen, but if you’re looking to build a voice assistant, Voxtral 3 is something you can actually use, rather than just something to look at.