Release 1.53.0

Anyone who dictates a lot knows the annoyance: you say the word you use all day long, and it comes out misspelled. Not once — every time, because the small model simply doesn't know that word. In a test with real speech, "gateway", said five times, came out as "ghetto" and "getway" all five.

A third engine, and the choice is yours

In Settings › Audio and Voice, next to System and Parakeet, comes Whisper. It gets the technical name and the slang right where the others miss, and the trade-off is written on the card itself, no fine print: it takes a few extra seconds after you release the key.

It is not a replacement. If you want the text instantly, stay on Parakeet, which is fast; if you speak jargon all day, pay the wait and get the text right. You can switch as many times as you like.

And the model now comes from our own house

A dictation model is the only piece the app needs to download in order to work, and it used to come from the service that publishes the model. On a day when that service was down, someone who had just installed the app would not be able to dictate.

Now the models live in the same place the app comes from. And if ours is down, it fetches from the usual one — the fallback stays, because swapping one single dependency for another single dependency would be no gain at all.