According to CoinMeta, Mistral has released Voxtral, which includes three speech transcription models, one speech synthesis model, as well as a complete set of voice, agent, and SDK. In the official demonstration, a user requested to make an appointment with a dermatologist in English, AI read a French referral letter, called the system to check for available times, and then confirmed the appointment in English after it was completed. The entire process was carried out through real-time voice dialogue. Voxtral adopts a modular design; the transcription models are responsible for recognizing speech, while general large models such as Mistral and small handle reasoning and invoking tools, and the speech synthesis model is responsible for providing responses. This SDK supports both telephone and web calls, allowing users to interrupt at any time, and Mistral claims that the delay for a single round of interaction is less than 1 second. The languages supported by transcription have expanded from 13 to over 50, and speech synthesis supports more than 35 languages and over 50 dialects. API charges $0.008 per minute for real-time transcription and $13 per million characters for speech synthesis.