Mistral has unveiled Voxtral 3, a comprehensive voice agent platform that brings together speech recognition, conversational reasoning, and expressive speech synthesis in a unified real-time pipeline. The release targets enterprise use cases including contact centers, virtual assistants, and hands-free copilots.
Four Pillars of Reliable Voice Agents
According to Mistral, building reliable voice agents requires solving four challenges simultaneously: accurate listening across languages and accents, deep contextual reasoning, natural emotional speech, and multi-step tool integration. Voxtral 3 addresses each of these with specialized models and a new agent SDK built on LiveKit.
The speech-to-text model, Voxtral Small Transcribe 3, achieves state-of-the-art results in multilingual transcription and diarization, outperforming Whisper, Scribe v2, and Gemini 3.5 Transcribe on benchmarks. It supports 50+ languages including 24 EU languages and extends to low-resource Asian and Indic languages. The Mini variant runs at 4B parameters with pricing at $0.0035 per minute.
Real-Time Pipeline Architecture
Voxtral 3 introduces a voice agent SDK that runs speech recognition, reasoning, and speech synthesis as a single streaming pipeline. During a call, Voxtral Mini Transcribe Realtime 3 listens and detects end-of-turn, a general LLM like Mistral Small 4 reasons and calls tools, and Voxtral Speech 3 replies with appropriate emotional tone. Each stage runs concurrently, achieving sub-second latency per turn and handling caller interruptions gracefully.
The system supports function calling, web search, and retrieval over enterprise systems during live conversations. Integration options include SIP for phone calls, WebRTC for browsers and apps, and connections to existing telephony and contact-center software.
Expressive TTS with Emotional Adaptation
Voxtral Mini TTS 3 generates speech that implicitly interprets conversational context to determine appropriate emotion, pace, and emphasis—without requiring explicit instructions. The model supports 35+ languages and 50+ dialects with voice cloning available from 30 minutes of audio through Pro Voice adaptation.
At $13 per million characters, Mistral claims Voxtral Speech 3 costs less than comparable offerings from OpenAI TTS-1, Deepgram, ElevenLabs Flash, and Gemini Flash.
Data Sovereignty Option
Like other Mistral models, Voxtral 3 supports both cloud API deployment (EU, US, and global endpoints) and self-deployment on customer infrastructure. This dual approach addresses enterprise requirements for data sovereignty and regulatory compliance, particularly in EU markets.
The voice agent API pricing is $0.04 per minute when using Mistral Small 4 for reasoning, or $0.08 per minute with other reasoning models.