Key points
- Mistral AI released Voxtral 3, a voice agent package with a transcription model, a text-to-speech model and an SDK built on LiveKit.
- Transcription handles more than 50 languages, while synthesized voices span over 35 languages plus regional dialects and can vary their emotional tone.
- Teams can self-host the stack or call EU-hosted APIs; transcription starts at $0.0035 per minute.
- Agents can reach outside systems mid-call, such as scheduling appointments or looking up invoices.
Why it matters for marketers: Marketers and support teams can prototype multilingual voice agents for bookings and billing questions with lower data-residency worries, and a published per-minute rate makes cost estimates simple.
Mistral AI has released Voxtral 3, a set of models and tooling for building voice agents that respond in real time. The package pairs a transcription model, Voxtral Small Transcribe 3, with a text-to-speech model, Voxtral Mini TTS 3, and an SDK built on LiveKit. According to The Decoder, these pieces run as one streaming pipeline that listens, reasons, calls tools and replies aloud.
The language coverage is broad. Transcription works in more than 50 languages, and the voice side covers over 35 languages along with regional dialects, adjusting its tone to fit the moment. During a conversation, an agent can connect to other systems, for instance to book an appointment or check an invoice.
For marketers and content teams, this lowers the barrier to voice channels such as phone-based lead qualification, customer service and multilingual brand assistants. Mistral says the stack can run on a company's own infrastructure or through EU-hosted APIs, which matters for teams handling customer recordings under strict data rules. Pricing for transcription starts at $0.0035 per minute, which gives planners a clear baseline for budgeting.
Worth watching: how natural the emotional voices sound in real customer calls, how well tool connections hold up under live conditions, and what the text-to-speech side costs, since the source only gives the transcription rate. Teams with a voice use case could run a small pilot against the documentation before committing.
Sources
This brief is our own summary. Read the full reporting at the source.
