Product
Words appear while you are still talking
Hold the shortcut and speak. Voicecape streams the audio as you talk and the sentence lands at your cursor — not after you stop, but while you are still finishing it.
Latency you can feel, measured rather than claimed
The number that decides whether dictation feels alive is the gap between you finishing a phrase and the text being final. Ours is 491ms at p95, measured end to end on the same synthesised audio we ran every candidate vendor through — the harness and the raw numbers are in the repository, not in a marketing claim.
That measurement is why recognition runs in the cloud. The model bundled inside the app answers in 699ms on the same audio and gets 15.8% of Korean characters wrong; the cloud model gets 4.4% wrong. Speed and accuracy are what this product sells, and they are not available on a laptop CPU today.
Five languages, and the sentences that mix them
English, Korean, Spanish, Japanese and Chinese, with no menu to switch. The hard case is not a Korean sentence or an English one — it is the sentence that is both, which is how anyone working in software actually talks. "useEffect 디펜던시 어레이" is one phrase in two scripts, and the bundled offline model gets 52.6% of that kind of sentence wrong while the cloud model stays near its single-language rate.
It is a switch, and the switch is real
Cloud dictation is a setting. Turn it off and recognition runs inside this Mac using the model that ships in the app — no audio leaves, and you can confirm that from the outside with airplane mode or a firewall monitor rather than by reading a policy. The Trust Center lists every host the code can reach and what goes to each one.
Recent changes here
Questions about this
Does it work while I am offline?
Yes — dictation falls back to the model bundled in the app. It is slower and less accurate, and it is there so a dropped connection does not stop you mid-sentence.
Is my audio kept anywhere?
No. Audio is streamed through our relay to the speech vendor and neither we nor they retain it. The Trust Center names the vendor and what reaches it.
Do I have to pick a language first?
No. The five supported languages are recognised together, including inside one sentence.