Less typing.
More thinking.

With cloud dictation on, the audio of that dictation goes to Soniox Inc., which turns it into text. It is processed in transit and not stored, under a data processing agreement we have signed. The text can then go to Cloudflare, Inc. to be cleaned up, with Google LLC, OpenAI, L.L.C. and Groq, Inc. behind it when the first is unavailable. Every one of those companies is named in the privacy policy, and the terms are the same whether you are on the trial or paying. With cloud dictation off, recognition runs on your Mac and none of that happens.

AI dictation tools usually rewrite what you said into “better” text. Voicecape does the opposite: it removes fillers and accidental repeats, and keeps your words, order and meaning exactly. If a sentence needs rewriting, that stays your decision.

Recognition falls back to your Mac and dictation keeps working — on a flight, in a dead zone, on a locked-down network. Accuracy is not identical between the two paths, and the figures on this page describe the cloud one.

English, Korean, Spanish, Japanese and Chinese. You do not pick one before you talk: leave it on automatic and Voicecape follows what you say, and a sentence that mixes two comes back mixed the same way, because the cleanup is instructed never to translate. On our benchmark set the character error rate is 4.4% for Korean and 2.3% for Japanese — synthesised audio, scored against the text a dictation product should produce.

The app stops inserting text. Your personal dictionary, your settings and anything you have already written stay exactly where they are, and starting a subscription picks up where you left off.