AI Lab · by Paolo Garcia · 2026-10-20
Mistral drops Voxtral 3, their next speech/audio model
mistral.ai
Mistral is back on the audio side with Voxtral 3. They've been quietly building a real alternative to the big US speech stacks, and a third generation suggests they're serious about it rather than treating audio as a side project. Curious whether this closes the gap on noisy, multilingual real-world audio or mostly improves benchmark numbers. Anyone already running it against Whisper-style pipelines or the closed APIs?
34 sparks · 18 comments
Comments
Oliver Smith · 12 likes
honestly at this point speech-to-text feels mostly solved for english. the interesting part is latency for voice agents
Minji Kim · 12 likes
latency and turn detection. nobody talks about turn detection and its the thing that makes voice bots feel dumb
Avery Johnson · 11 likes
my voice agent still interrupts me when I take a breath so yes please
Curious River · 9 likes
European speech model that isn't a whisper finetune, I'm here for it
Pixel Panda · 0 likes
the eu ai scene needed a win honestly
Can Arslan · 6 likes
Is the license the same as Voxtral 2 or did they change it? That decides whether we can even touch it at work
Ethan Brooks · 8 likes
didn't see anything obvious yet, would wait for the model card before getting excited
Lucía García · 2 likes
finally. been waiting to replace our transcription vendor with something we can actually self host
Mango Rider · 8 likes
same, the per-minute pricing on the closed apis adds up stupid fast
Diego Quispe · 1 likes
the real test for me is accented french and code switching. every model looks great on clean podcast audio
Pixel Panda · 12 likes
would love someone to post a proper eval on that instead of the usual leaderboard screenshots
Can Arslan · 0 likes
this. my parents switch between arabic and french mid sentence and everything falls apart
Jonas Schneider · 1 likes
anyone know if it runs reasonably on a single consumer gpu or is this a datacenter thing
Valeria López · 0 likes
Mistral naming is so much better than everyone else. Voxtral, Codestral, Magistral. meanwhile others are on model-5.5-pro-max-preview
Lucas Dubois · 0 likes
Voxtral 3 Ultra Thinking Turbo coming q1 probably
Maya Putri · 0 likes
tried the previous version for meeting notes and it was fine but diarization was the weak spot. hoping that improved
Mariam Said · 2 likes
diarization is always the weak spot lol, its like the final boss of audio