Features
What Vocelle does, in detail
Everything here ladders up to one of three pillars: local by architecture, types anywhere a cursor blinks, built like infrastructure. If a feature doesn't serve one of those, we didn't build it.
Local by architecture
Vocelle transcribes with Whisper large-v3-turbo, running entirely on your Mac's Neural Engine. Your audio is held in memory only for the moment it's being turned into text, then it's gone. It is never written to disk, and it never has anywhere to go — there is no server in the transcription loop, so there's no server to trust.
- Whisper large-v3-turbo runs 100% on-device, on the Neural Engine.
- Audio lives in memory only. It is never saved to disk, ever.
- Works with no internet connection. There's nothing to connect to.
- The only thing Vocelle ever downloads from us is the speech model itself, once.
Types anywhere a cursor blinks
Hold right-⌘ from anywhere on macOS, speak, and release. Vocelle synthesizes real keystrokes at your cursor, so the text lands exactly where you're working — including places clipboard-based tools and browser extensions can't reach.
- One shortcut, system-wide: hold right-⌘, speak, release.
- Types into terminals, code editors, native apps, and browsers alike.
- A live waveform overlay shows Vocelle is listening while you speak.
- Hands-free mode, for when holding a key isn't convenient.
Built like infrastructure
Depth you can verify, not adjectives you have to believe. Every dictation is written to a crash-safe log before Vocelle processes it, so a crash mid-sentence costs you nothing. An on-device model then cleans the transcript up — and a guardrail keeps that cleanup honest.
- A crash-safe write-ahead log: your words are saved before they're processed, so you never lose one.
- Searchable history of everything you've dictated, with one-click delete.
- A personal dictionary for names and terms Whisper wouldn't otherwise know, with one-click import from other dictation apps.
- Fully offline, start to finish.
The guardrail, told plainly
What happens when cleanup goes wrong
After Whisper transcribes your speech, an on-device language model tidies the result: it drops filler words, fixes punctuation, and resolves the times you correct yourself mid-sentence. That model is not allowed to do anything else.
Vocelle checks the cleaned-up text against what you actually said. If the result keeps at least half of your original words, it types. If the model strays past that line — drops too much, adds things you didn't say, starts answering your question instead of transcribing it — Vocelle throws that output away and falls back to deterministic cleanup: filler removed, punctuation fixed, nothing rewritten, nothing invented.
In short: the cleanup model can tidy your words. It can never replace them. That's not a promise — it's a rule enforced on every dictation, and we're telling you about it in public because it's a failure mode worth documenting, not hiding.
See the honesty section
A plain-language table of exactly what Vocelle can and cannot see. No hedging, no legalese.
Read what Vocelle can and cannot see