Mistral Unveils Voxtral 3 — Voice Models and SDK for Voice Agents

Mistral AI has introduced Voxtral 3, a family of speech recognition and synthesis models featuring a LiveKit-based Voice Agent SDK that integrates listening, LLM reasoning, tool calling, and speech output into a single streaming pipeline with sub-second latency.

Compiled by Sergey KostenchukPublished 2026-10-20Updated 2026-10-20

2026-10-20 Coding OpenAI

🗒 Mistral Unveils Voxtral 3 and Voice Agent SDK for Voice Agents

The new family of speech synthesis (TTS) and recognition (STT) models, paired with a LiveKit-based SDK, connects listening, LLM reasoning, tool calling, and speech output into a single streaming pipeline with sub-second latency. The claimed WER is 4.0–4.1% on FLEURS 45, compared to 14.9% for Whisper large-v3.

🌍 Open weights and self-deployment are replacing stitched-together 'STT + LLM + TTS' setups: contact centers, telecom, and fintech are reducing costs and risks associated with transmitting audio to third parties. Transcription starts at $0.0035/minute, and TTS is $13 per 1 million characters.

👤 The SDK, supporting SIP and WebRTC, eliminates the need for manually assembling Whisper + LLM + TTS. The figures are vendor-reported; verify them on your own data. Voice cloning using 30 minutes of audio without consent mechanisms carries impersonation risks.

Source 1: https://mistral.ai/news/voxtral-3/ Source 2: https://cryptobriefing.com/mistral-unveils-voxtral-3-voice-agent-sdk/

Sources