Mistral AI unveils voxtral 3 with speech-to-text, tts and voice agent sdk

Mistral AI has launched Voxtral 3, a speech model family paired with a voice agent SDK. The release includes three transcription models, one text-to-speech model, and an SDK built on LiveKit. All four models are in public preview via the Mistral API. The transcription models support diarization and context biasing, with batch transcription covering 50 languages, including 24 official EU languages. The SDK operates as a streaming pipeline with three stages, enabling real-time transcription, reasoning, and speech generation. Voxtral Speech 3 replies in the tone of the conversation, with each turn completing in under one second. Voxtral 3's performance metrics include a 2.1% AA-WER for the Small model, placing it second behind Fun-Realtime-ASR-preview. The realtime model on FLEURS 35 scores 5.7% WER at 2,400 ms, with lower accuracy at shorter latencies. Voxtral Speech 3 costs $13 per 1M characters, undercutting competitors like OpenAI TTS-1 and ElevenLabs Flash. The model offers explicit controls for emotion, pacing, and emphasis, with Pro Voice adapting a speaker’s style from 30 minutes of audio. Mistral’s models are available on European, US, and global endpoints, with OpenRouter also listing them.