Everything under Settings → Transcription controls how audio becomes text. All transcription runs on-device. Nothing is sent to the cloud.
Language
talat uses one of two on-device speech models, and this is where you choose between them:
- English only. Tuned for English, so it’s a little faster and a little sharper on English speech.
- A mix of languages. A multilingual model that spans dozens of languages, from English to Japanese, Arabic, or Korean.
You make this choice during onboarding and can change it here at any time. Switching downloads the other model if it isn’t already on your machine, and the change applies to your next meeting.
How transcription works
A real-time draft appears as you speak, then a final, higher-quality pass cleans it up once each phrase finishes. This happens on every recording; there’s nothing to turn on.
Vocabulary boost
Add words and phrases talat should listen for more carefully: proper nouns, brand names, and domain terms the model wouldn’t otherwise expect. Turning the boost off can make transcription slightly faster; changes take effect from your next meeting.
Telling speakers apart
Working out who said what is a separate feature that’s always on. How talat names people and recognises them across meetings is covered on the identify speakers page, and its cross-meeting behaviour is set under Settings → Speakers.