Mistral’s October 20 release is easier to explain by the invoice it replaces. Voxtral 3 is transcription, a speaking voice, and the glue that turns those into an agent that can take a phone call.
The pieces
There are three transcription models, including a smaller one and a realtime one, covering more than 50 languages. Voxtral Speech 3 handles text to speech in more than 35 languages and dozens of dialects. Mistral’s Small Transcribe model landed near the top of Artificial Analysis’s speech-to-text board, in third, and the speech model entered that site’s top twelve for synthesis.
The useful bits are the ones call centers already pay extra for. The transcriber can label who is talking. You can hand it names and jargon so it stops guessing at the words your company actually uses. The SDK, built with LiveKit, pipes realtime audio into a reasoning model that can call tools, then out through the speech model.
Who it is for
This is a developer product. Nobody downloaded a Mistral app and started dictating. A team that currently buys recognition from one vendor, a model from a second, and a voice from a third can now put that chain on one API. Mistral priced transcription under half a cent a minute, which is how these decisions actually get made.
Listen to it on your own audio before you cancel the old contract. Accents, hold music, and product names are where a leaderboard goes to die.