AI BriefingMistral AI releases Voxtral 3 voice agent SDK and speech models with sub-1-second turn times
·2026.10.20 21:00
Key point
The release includes a LiveKit-based Voice Agent SDK and speech models priced from $0.0035 per minute for transcription and $13 per 1M characters for TTS.
Details
- Mistral AI launched Voxtral 3, a suite of speech models and a Voice Agent SDK designed for real-time, multi-turn voice interactions with low latency and data sovereignty options.
- The Voice Agent SDK integrates listening, reasoning, tool calling, and speech into a single streaming pipeline, achieving turn completion times of under 1 second and supporting function calling, web search, and system integration.
- Voxtral Small Transcribe 3 supports 50+ languages and outperforms competitors in multilingual transcription and speaker diarization, achieving a 4.0% WER on the FLEURS 45 benchmark and 18.4% cpWER on diarization tasks.
- Voxtral Speech 3 (TTS) supports 35+ languages and 50+ dialects, priced at $13 per 1M characters, and offers adaptive control over emotion, pacing, and emphasis based on conversation context.
- Deployment is available via Mistral's EU-hosted API or self-hosted infrastructure, allowing control over audio processing location and data retention.
- Pricing includes $0.0035/min for Voxtral Mini Transcribe 3 (batch), $0.008/min for Voxtral Mini Transcribe Realtime 3, and $0.0040/min for Voxtral Small Transcribe 3.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.