Mistral releases Voxtral 3 speech model series, launches voice agent SDK

PANews reported on October 21, according to official news from Mistral, the French AI company Mistral has released the Voxtral 3 series of speech models, including the speech-to-text models Voxtral Small Transcribe 3 and Voxtral Mini Transcribe 3, the real-time transcription model Voxtral Mini Transcribe Realtime 3, and the text-to-speech model Voxtral Speech 3. It also launched a voice agent SDK based on LiveKit, supporting speaker diarization, voice cloning, function calling, and SIP telephony access. The transcription models support more than 50 languages, with API pricing ranging from $0.0035 to $0.008 per minute, while text-to-speech is priced at $13 per 1 million characters. Mistral stated that Voxtral Small Transcribe 3 outperforms models such as Whisper large-v3 and Gemini 3.5 Transcribe on the FLEURS benchmark.