Mistral AI has introduced Voxtral 3, a comprehensive language model family designed for real-time voice agents. The system supports over 50 languages for transcription and 35+ for speech output, with emotional adaptation.
TL;DR
- Mistral AI launches Voxtral 3, a real-time voice agent stack with speech-to-text, text-to-speech, and agent SDK capabilities.
- The system supports 50+ languages for transcription and 35+ for speech output, with emotional adaptation and tool use during live conversations.
- Voxtral 3 can be deployed on users' own infrastructure or through EU-hosted APIs, with pricing starting at $0.0035 per minute for transcription.
What happened
Mistral AI has unveiled Voxtral 3, a new family of language models tailored for real-time voice agents. The system integrates speech-to-text, text-to-speech, and a voice agent SDK based on LiveKit, enabling seamless listening, thinking, tool use, and speaking in a single streaming pipeline.
Voxtral 3 includes the Voxtral Small Transcribe 3 model for speech-to-text, supporting over 50 languages, and the Voxtral Mini TTS 3 model for text-to-speech, covering over 35 languages and 50 dialects with emotional adaptation. The system also allows agents to interact with external systems during live conversations, such as booking appointments or checking invoices.
Why it matters
For developers and startups, Voxtral 3 offers a comprehensive toolkit to build advanced voice agents with multilingual capabilities and real-time processing. The ability to deploy the system on users' own infrastructure or through EU-hosted APIs ensures data control and privacy, which is crucial for enterprises handling sensitive information.
Investors may see this as a strategic move by Mistral AI to expand its product offerings and capture a larger share of the growing voice AI market. The competitive edge lies in the system's real-time processing and multilingual support, which can cater to a global audience.
Key facts
- Voxtral 3 supports over 50 languages for transcription.
- Text-to-speech output covers over 35 languages and 50 dialects with emotional adaptation.
- The system includes a speech-to-text model (Voxtral Small Transcribe 3), a text-to-speech model (Voxtral Mini TTS 3), and a voice agent SDK based on LiveKit.
- Agents can interact with external systems during live conversations.
- Voxtral 3 can be deployed on users' own infrastructure or through EU-hosted APIs.
- Pricing starts at $0.0035 per minute for transcription.
- More details are available in the official documentation.
Context
The launch of Voxtral 3 positions Mistral AI as a significant player in the voice AI market, competing with other major players like Google, Amazon, and Microsoft. The demand for real-time voice agents is growing, driven by applications in customer service, healthcare, and personal assistants.
As AI technologies continue to evolve, the ability to process and generate human-like speech in real-time is becoming increasingly important. Mistral AI's Voxtral 3 aims to meet this demand with a comprehensive and scalable solution.
