Live Typing
Words land in the field while you're still speaking — underlined and provisional, committed as one clean insert. One ⌘Z undoes the whole utterance. No other dictation app ships this.
Dictation that never
leaves your Mac.
Hold a key. Speak. Release. Your words appear at the cursor, cleaned up — transcribed and polished entirely on this machine.
Hold the key to dictate — or just watch.
Free & open source · Apple silicon · macOS 26 (Tahoe) or later · SHA-256 · Notarized by Apple
The whole pipeline fits in the room.
Every stage runs on your Mac's own silicon. There is no networking code on any path — not disabled, not opted out. Absent.
- Microphone records only while the key is held
- Speech engine Apple SpeechAnalyzer, on the Neural Engine — words stream in while you speak
- Cleanup cage on-device Apple Intelligence fixes misheard words, strips filler — verbatim always wins
- Your rules personal dictionary and voice commands, guaranteed by regex, not a model
- Your cursor 69–108 ms from key-release to final text
$ network requests made while you dictated 0 — and 0 yesterday, and 0 forever. Audio and text stay here.
Built like a tool, not a service.
AI cleanup, caged
The on-device model fixes homophones ("right heavy" → "write-heavy") and punctuation, inside strict guard rails: if it rewrites, summarizes, or answers instead of cleaning, you get your verbatim words back. Cleanup can win, never lose.
Spell it once, keep it forever
"Add Sharique to the invite — that's S-H-A-R-I-Q-U-E." The misheard word is fixed in place and the spelling joins your dictionary, so recognition itself improves next time.
Your dictionary, enforced
Names, jargon, and product terms are corrected by deterministic rules on every utterance and fed to the recognizer as biasing hints. Your vocabulary is a guarantee, not a suggestion.
Privacy by construction
No account, no subscription, no telemetry, no network calls anywhere in the binary. History is plain text in a local file you can purge from the menu. The mic is hard-off between holds.
Fast where it counts
The speech engine pre-warms on key-down. Partial words appear about a second into the hold, and the finished, cleaned text lands 69–108 ms after release.
Where your voice goes.
| Wisprit | Wispr Flow | macOS Dictation | |
|---|---|---|---|
| Speech processed | On your Mac | Cloud (AWS) | On your Mac |
| AI cleanup | On-device, caged | Cloud LLM | None |
| Works offline | Always | No | Yes |
| Account & price | None · free | Account · $15/mo | None · built-in |
| Custom vocabulary | Yes, learns & biases | Yes, cloud-synced | No |
| Live text while speaking | In the field itself | No | Yes |
| Your exact words | Verbatim-first, always recoverable | Rewrites by default | Verbatim only |
Wispr Flow is excellent software with a different architecture: your audio, screen context, and history are processed on its servers. Wisprit exists for people who want the same class of experience with none of that leaving the room.
What you need
- An Apple silicon Mac on macOS 26 (Tahoe) or later — the on-device speech engine ships with the OS.
- Apple Intelligence enabled for AI cleanup. Without it, dictation still runs with deterministic cleanup only.
- Three permissions, granted once: Microphone, Accessibility, Input Monitoring. The built-in
doctorcommand walks you through every one.