Two new speech-to-text models (similar to Whisper) from Mistral today - one of them is API-only, the other is a 8.9GB Apache-2.0 licensed open weights model for "realtime" transcription. They're both very good! https://simonwillison.net/2026/Feb/4/voxtral-2/
Discussion
Loading...
@simon
It's the first one you can run yourself that does diarization, isn't it? I've seen hacks to implement it that were painful to use before, but nothing truly integrated.
![Mat]3](https://media.metalhead.club/accounts/avatars/115/712/242/382/786/036/original/c4d299b8806545be.jpg "Mat]3")
@simon Whisper is very good, but when the audio is noisy and not very clear, it starts hallucinating. I wonder how Voxtral fares in this case. The first Voxtral was just a bit worse than Whisper.
