Product
The network drops, the sentence does not
A 547MB speech model ships inside the app — nothing downloads on first run. When there is no network, recognition moves to it and you keep dictating.
A feature, not the product
We are direct about what this is. The bundled model answers in 699ms against the cloud’s 491ms, and gets 15.8% of Korean characters wrong against 4.4% — on Korean-English mixed sentences the gap is far wider. It is there so a dropped connection, a plane or a conference wifi does not stop you. It is not what this product competes on.
It is also the answer to a different question
Turning Cloud dictation off makes the offline model the only recogniser, permanently. That is a legitimate way to use Voicecape if the audio must not leave the machine, and it is a setting rather than a separate build — you can verify the difference from the outside with airplane mode.
Questions about this
Do I have to download the model?
No. It ships inside the app.
How much disk does it use?
About 547MB, inside the app bundle.
Can I use only the offline model?
Yes — turn Cloud dictation off and it becomes the only recogniser.