Languages and models

What it translates into

21 languages, and the list works both ways — a language you can translate into is also a language you can translate from. One direction ships inside the app: ru/uk→en. Every other direction is a dictionary you download when you pick the language, and that includes the reverse en→ru/uk.

Only the ru/uk→en dictionary ships inside the app. Every other direction, including the reverse en→ru/uk, is downloaded from cdn.verbaide.com when you ask for it.

What it hears

Speech recognition runs on a downloaded model, not on Apple’s engine — Apple’s own Russian dictation does not add punctuation, and without punctuation both the translation and the readability fall apart.

Apple’s own engine covers twelve languages, English among them. Its models are downloaded too — Settings → Models has a tab for them.

Some languages can be read but not heard: Serbian, for instance, is a language the app can translate for you but not one it can listen to.

What gets downloaded, and when

The main speech model. Pulled the first time it is needed — usually on the try-it screen of the first-run setup. Nothing works without it, including free dictation.

A term-dictionary model (98 MB). Installed in the background on your first dictation. This is what makes “Гитхаб экшн” come out as GitHub Actions.

Translation dictionaries. One per direction, downloaded when you select the language.

The local style model. Only if you choose to install it, and only for the paid style presets. The weights come from Hugging Face; the app downloads them itself.

Where to see all of it

Settings → Models. Each tab shows what is installed, what is downloading and what is available: recognition, translation dictionaries, Apple’s recognition, and the local AI.

How much disk it takes

That depends on how many languages you picked and whether you installed the local style model, so there is no single honest number to print here. Settings → Models shows the size of each item before you download it.