Mistral announces Voxtral 3 speech models, including Turkish, and a voice agent toolkit

Mistral AI announced the Voxtral 3 speech model family on 20 October 2026. According to the company's page, the release also includes a voice agent SDK built on LiveKit that runs speech-to-text, reasoning, tool calling and speech generation as one streaming pipeline. Agents can connect to phone lines over SIP and to browsers and apps over WebRTC.

• Voxtral Small Transcribe 3 handles transcription and speaker diarization, and the 4-billion-parameter Voxtral Mini Transcribe 3 is offered for batch transcription.
• Voxtral Mini Transcribe Realtime 3 targets latency under 500 milliseconds in more than 35 languages; Voxtral Mini TTS 3 provides voice cloning and streaming speech.
• Transcription covers more than 50 languages; Turkish appears in the transcription evaluation list, the realtime model's language list and the speech sample library.
• Prices start at 0.0035 dollars per minute; speech generation costs 13 dollars per million characters.

Why it matters

Voice-driven work such as call centers, booking lines and meeting recordings needs models that understand Turkish well and respond with low latency. With Turkish in its language lists and direct phone-line connectivity, Voxtral 3 could give companies another option for building voice assistants. Mistral offers European, US and global API endpoints as well as self-deployment; since call recordings contain personal data, companies would do well to decide early which option fits their KVKK obligations. The performance comparisons are published by Mistral itself, and some are labelled as internal evaluations.