yap

dictation for macos. hold fn, talk, let go, and it's pasted wherever your cursor is.
it uses the speech model that comes with macos, so nothing leaves your mac and it's free.

i wrote it to replace wispr flow, which sat on 1.3 gb of ram and about 5% cpu whether i was talking or not. yap sits on about 0.1 gb and only does anything while you're talking.

install

brew tap 0xdeafcafe/yap https://github.com/0xdeafcafe/yap
brew install --cask yap

it'll want the microphone (to hear you) and accessibility (to see fn and press ⌘V). quit wispr flow first, or both will paste.

use

hold fn	talk, let go to paste
double-tap fn	hands free, tap again to paste
fn + another key	cancel, so fn+← still works
esc	cancel

or click the blob to start and click it again to paste. hold ⌘ and drag it to move it to another edge.

yap takes fn over while it's running, so the emoji picker and apple's dictation don't open. your setting comes back when you quit.

the menu bar icon has your last few transcripts and opens settings (⌘,), where you'll find your history, the language you dictate in (any apple's speech model knows, from british english to brazilian portuguese), a few app icons, and opt-in formatting ("new line", "bullet", "code npm test end code").

it learns your words. when you start talking it reads the window you're in and takes the names and odd words from it (anything with capitals in the middle, digits, or that the dictionary doesn't know), plus whatever an app like rush sends it, and tells the speech model to listen for them. if it then hears something that's one letter or so off exactly one of them, like "reddis" with "Redis" on screen, it fixes it, and after a few times it learns the swap for good. it never touches an ordinary word, and it doesn't keep what it read. you can also fix a name after it pastes, or in history, and it'll spell it right next time. words you keep deleting, like "basically", get dropped. settings › words shows what it's learned and suggests names from your history that the dictionary doesn't know. you can add your own to ~/.config/yap/words.txt too, one per line, btw => by the way to swap a phrase, or drop: basically to drop a filler.

privacy: transcripts are kept for 30 days, then deleted. if you already had older history, yap tells you how much would go first and lets you keep it forever. recordings are off unless you turn them on. settings › privacy shows how much is kept and lets you stop saving what you say, change how long things are kept, or delete the lot. nothing leaves your mac.

in other apps

yap says what it hears, as it hears it, on ~/.config/yap/live.sock: one line of json per change, with the words it's sure of, the ones it's still guessing, how loud you are, and the app it'll paste into. only your user can connect. rush uses it to show the dictation in its message box while you talk.

an app can tell yap which words to listen for by writing a line back, when it connects and whenever its words change:

{"context":{"app":"com.mitchellh.ghostty","words":["Redis","rushd","LangWatch"]}}

app is the bundle id of the terminal it runs in, or "" if it doesn't know. yap uses the words when it's pasting into that app (or any app, for ""), keeps only the latest line from each connection, and forgets it when the connection closes. up to 200 words; lines over 64 kib and anything else are ignored.

is it any good

i ran 40 of my old wispr flow dictations (about 10 minutes) through each engine on an m1 max.

engine	error rate	wait after you stop	ram
parakeet v3 (mlx)	7.4%	0.5 s	1.1 gb
wispr flow (cloud)	7.4%	0.6 s	1.3 gb, all the time
apple speechtranscriber (yap)	8.6%	0.9 s	0.1 gb
whisper large-v3-turbo	11.1%	1.4 s	0.9 gb

nobody hand-checked these, so each one's scored against what the others agreed on, which is fine for comparing them but isn't an exact accuracy.

wispr flow cleans up filler with an llm. apple's on-device model deleted real words when i tried it ("i really don't like", "looks terrible"), and qwen3 4b wanted 2.9 gb and another 1.5 s, so yap uses plain rules: um, uh, repeated words, and a comma'd "like" or "you know". none of them took out a real word on the same 40 clips. you can turn the on-device model on in settings if you want its grammar fixes, but yap only uses its answer if every word it took out was a filler.

build

needs macos 26 and xcode's command line tools.

./bundle.sh && open Yap.app
Yap.app/Contents/MacOS/Yap --selftest     # the rules and the tests
Yap.app/Contents/MacOS/Yap --demo bottom  # play the panel on the bottom edge
Yap.app/Contents/MacOS/Yap --settings     # open the settings window on its own

every dictation is logged to ~/Library/Logs/Yap/dictations.jsonl, for as long as settings › privacy says (30 days unless you change it). audio's only kept if you turn it on there.

to release, bump the versions in Info.plist and push a v<version> tag. the workflow signs, notarises and publishes it.

it leans on two private apis, one to stop fn opening the emoji picker and one for the hand cursor. both need replacing before it could go on the app store.