Mistral unveils Voxtral 3 with advanced voice agent SDK and speech models

LiveKit: LiveKit supplies real-time communication infrastructure for audio and video applications. It forms the foundation of the Voxtral 3 voice agent SDK, enabling the unified streaming pipeline that connects transcription, model reasoning, and speech synthesis. The integration supports telephony, WebRTC, and existing contact center systems. Mistral: Mistral AI develops advanced open-weight large language models focused on sovereign European AI infrastructure and enterprise solutions. The company recently introduced Voxtral 3 to expand its platform with specialized speech models and a voice agent SDK. It offers EU-hosted API access and self-deployment options to support data control in sensitive applications. Voxtral: Voxtral is Mistral's family of speech-to-text and text-to-speech models built for multilingual voice AI workflows. Voxtral 3 marks the latest iteration, adding a voice agent SDK on LiveKit for integrated real-time transcription, reasoning, and expressive speech output. The lineup supports professional adaptations and deployment across diverse languages and dialects. Pro Voice: Pro Voice provides advanced voice adaptation capabilities within the Voxtral ecosystem for high-fidelity replication of individual speakers. Launched alongside Voxtral 3, it targets professional scenarios requiring consistent emotional range and style across multiple languages. Users supply sample audio to create custom voices suitable for branded interactions. Voxtral Speech 3: Voxtral Speech 3 functions as the expressive text-to-speech component in Mistral's Voxtral lineup, supporting instruction-based and implicit emotional control. It powers the speech output stage of the new voice agent SDK announced in Voxtral 3. The model emphasizes natural prosody and dialect accuracy across supported languages. Voxtral Mini TTS 3: Voxtral Mini TTS 3 delivers state-of-the-art expressive text-to-speech synthesis with strong instruction following. It appears in the Voxtral 3 announcement as the core speech component for natural, context-aware voice responses across languages. The model adapts emotion, pace, and emphasis based on conversational cues or explicit tags. Voxtral Small Transcribe 3: Voxtral Small Transcribe 3 serves as Mistral's leading model for accurate multilingual speech transcription and speaker diarization. Featured prominently in the Voxtral 3 release, it incorporates context biasing and end-of-turn detection to support responsive voice agents. The model handles domain-specific vocabulary without additional fine-tuning. Voxtral Mini Transcribe Realtime 3: Voxtral Mini Transcribe Realtime 3 is a streaming transcription model optimized for low-latency agentic voice applications. Included in the Voxtral 3 family, it combines diarization and context biasing with rapid end-of-turn prediction. The model enables seamless integration into real-time pipelines for live conversations. Integration: Voice agent platforms increasingly combine specialized speech models with general reasoning models through unified real-time SDKs built on established communication frameworks. Sovereignty: Mistral prioritizes European data sovereignty by providing EU-hosted API endpoints and self-deployment options for voice processing workloads. Multilingual Reach: Recent speech model releases emphasize expanded coverage of languages and regional dialects to support global conversational AI deployments.