Mistral AI Voxtral 3 release notes

Overview

Mistral AI has introduced Voxtral 3, a new family of speech models and a voice agent SDK designed to facilitate real-time, multi-step conversational AI. The system integrates transcription, reasoning, and speech synthesis into a single streaming pipeline, supporting over 50 languages and enabling deployment via Mistral’s EU-hosted API or on private infrastructure.

Speech-to-Text Capabilities

The Voxtral 3 transcription models are engineered for high-accuracy multilingual processing and real-time agentic workflows. Key technical features include:

  • Voxtral Small Transcribe 3: Optimized for high-accuracy multilingual transcription and diarization, supporting 50+ languages. It features context-biasing, allowing developers to supply domain-specific vocabulary and keyphrases to improve recall without fine-tuning.
  • Voxtral Mini Transcribe Realtime 3: Designed for low-latency agentic use cases, achieving sub-500ms latency. It includes native end-of-turn detection that utilizes both acoustic and semantic signals to minimize false cut-offs during user interruptions.
  • Performance: The models are reported to outperform existing benchmarks like Whisper and Scribe v2, with the 4B parameter Mini variant priced at $0.0035/min for batch processing.

Text-to-Speech and Expressive Synthesis

Voxtral Mini TTS 3.0 focuses on natural, context-aware speech generation with support for 35+ languages and 50+ dialects.

  • Emotional Adaptability: The model implicitly interprets semantic context to adjust prosody, rhythm, pacing, and emphasis without requiring explicit instruction.
  • Pro Voice Adaptation: For professional use cases, the system can reproduce a specific speaker's emotional range and character using 30 minutes of audio.
  • Instruction Following: Users can control speech output via variables such as emotion, pacing, and emphasis.
  • Cost Efficiency: The model is priced at $13 per 1 million characters, positioning it as a competitive alternative to existing enterprise TTS solutions.

Voice Agent SDK

The Voxtral 3 voice agent SDK provides a unified streaming pipeline that combines listening, reasoning, and speaking. Built on LiveKit, the architecture allows for modularity, enabling developers to swap the underlying LLM, voice, or language without reconfiguring the entire pipeline.

  • Integration: The SDK supports SIP, WebRTC, and standard telephony/IVR systems, allowing agents to perform multi-step tasks like identity verification, billing inquiries, and appointment scheduling.
  • Deployment: Organizations can maintain data sovereignty by deploying the full stack on their own infrastructure or utilizing Mistral’s EU-hosted API endpoints.

Pricing and Availability

Voxtral 3 is available immediately via the Mistral API. Pricing is structured as follows:

Model Price
Voxtral Mini Transcribe 3 (batch) $0.0035/min
Voxtral Mini Transcribe Realtime 3 $0.008/min
Voxtral Small Transcribe 3 $0.0040/min
Voxtral Speech 3 (batch/realtime) $13/1M chars

Sources