French AI Firm Mistral Releases Voxtral 3: Integrating Transcription, TTS, and Voice Agent SDK
Mistral launches the Voxtral 3 speech model family and SDK, supporting 50+ language transcription and emotion-adaptive TTS, with the Small model ranking third on independent benchmarks.
ImportanceMaterialEvidenceE3 inspectableWrite-upStandard
French artificial intelligence company Mistral AI officially released Voxtral 3 on October 20, 2026. This product family includes three speech-to-text (ASR) models, one text-to-speech (TTS) model, and a voice agent SDK built on LiveKit. The release aims to address challenges in end-to-end voice agents regarding real-time performance, multilingual support, and complex workflow handling.
Voxtral Small Transcribe 3 is the flagship batch transcription model, supporting 50 languages (including 24 official EU languages) with speaker diarization and context biasing capabilities. According to data from independent evaluation firm Artificial Analysis, its weighted average word error rate (AA-WER) is 2.1%, ranking third on the platform behind Fun-Realtime-ASR-preview and Microsoft's MAI-Transcribe-2. Mistral's self-reported data shows a forced-language WER of 4.1% on the FLEURS 45 benchmark, outperforming Whisper large-v3 (16.0%).
Beyond transcription, Voxtral Mini TTS 3 introduces natural speech synthesis technology capable of adjusting emotion, pace, and emphasis based on conversational context, supporting over 35 languages and 50+ dialects. The accompanying Voice Agent SDK integrates real-time transcription, large language model reasoning (such as Mistral Small 4), and speech generation into a single streaming pipeline. Each turn completes in under one second, with support for phone systems via SIP and browser/app sessions via WebRTC.
All models are currently in public preview and accessible via the Mistral API. Pricing for Voxtral Mini Transcribe 3 batch transcription is $0.0035 per minute, Small version is $0.004 per minute, realtime transcription is $0.008 per minute, and TTS services cost $13 per million characters. Notably, while batch transcription performs strongly, Voxtral Mini Transcribe Realtime 3 has a word error rate of 7.4% at low latency (240 ms), higher than Gemini 3.5 Realtime's 4.2%, indicating an accuracy trade-off in extreme low-latency scenarios.