yap

Dictation and meeting transcription for macOS, all of it on your Mac.

Hold a key, talk, let go. The text turns up at your cursor about 60 ms later. No account, no upload. The network gets used once, to fetch the model.

Install

brew install --cask terrifiedbug/tap/yap
yap setup

yap setup asks for the permissions and pulls the model down. That is 220 MB, once, and then you are offline forever.

Builds are signed with a Developer ID certificate and notarized by Apple, so there is no Gatekeeper prompt. Apple Silicon only, because the model runs on the Neural Engine.

Use

Hold fn, speak, release. A small pill shows up while the mic is live, and the same mark sits in the menu bar the whole time yap is running.

yap install --launch-at-login   # menu bar, back after every login
yap                             # foreground instead, dies with the terminal
yap record                      # record a meeting now, ^C to stop
yap doctor                      # permissions, key mapping, model

Dictation and the meeting recorder share one process and one loaded model. The recorder takes your mic and the system audio as two separate tracks and gives you one transcript, with timings and speaker labels.

If fn does something else on your Mac, yap doctor says how to get it back. There is also --hotkey, and dictation.hotkey in the config file.

"Copy last transcript" in the menu bar is there for the press that landed in the wrong window. yap holds the most recent one in memory and nowhere else.

"Quit yap" stops the background daemon until your next login. yap start brings it back sooner.

If you run a menu bar manager — Ice, Thaw, Bartender — and the mark is nowhere to be seen, look in its hidden section. Managers that file newly-appeared items there catch yap the first time it shows up. Reveal that section, then hold Command and drag the mark out of it once; it stays where you put it.

yap run	The daemon, in the foreground. The default.
yap start / yap stop	Start or stop the background daemon.
yap record	Record one session now, then transcribe it.
yap models list	The models, and which ones you have.
yap models download <id>	Fetch one early.
yap doctor	Permissions, key mapping, model, login item.
yap setup	Permissions and the model, in one go.
yap install	Add or remove the login item.
yap bench --audio FILE	Time it on your own audio.
yap --version	Print the version.

Configuration

~/.config/yap/config.json. Every key is optional and a flag beats the file. "Edit config…" in the menu bar opens it, filled in with the defaults.

{
  "recordings_dir": "~/Recordings",
  "meeting_detection": false,
  "mic_voice_processing": true,
  "on_stop": "my-hook",
  "transcription": { "enabled": true },
  "dictation": {
    "model": "parakeet-tdt-ctc-110m",
    "hotkey": "fn",
    "overlay": true,
    "newline_after_release": false,
    "mute_output": false
  }
}

Save it and yap picks it up. The hotkey, the overlay, mute_output, newline_after_release and meeting_detection all change on the spot. A new model or recordings_dir wants a restart, and yap says so when it sees one.

newline_after_release hits Return once the text is in, which is what you want for chat boxes.

mute_output silences the speakers while the key is down. Your mic hears the room and the room includes whatever you are playing, so a video behind a press gets transcribed along with you. Off by default because you can hear it happen.

meeting_detection offers to record when something else grabs the mic. Off by default, and nothing is watching until you turn it on.

mic_voice_processing cancels speaker echo on the mic track. On by default: a call coming out of your speakers goes back into the mic. Without it, the other side gets transcribed twice, the second time as you. If some audio route returns silence instead, yap notices inside a second and restarts the mic raw.

on_stop runs a command after each recording with the session folder as its argument.

transcription is automatic transcription of recordings, on by default. Dictation ignores it, since the hotkey always transcribes. When one finishes while the daemon is running, a banner drops under the menu bar with an Open button that reveals the transcript in Finder. Turn it off to use yap as a plain recorder: on_stop then fires when the recording stops rather than after the transcript. Nothing is lost either way. Turn it back on, restart, and yap works through every session under recordings_dir that has no transcript yet, firing on_stop again for each. Anything you put somewhere else with yap record --out is left alone.

Models

id	languages	size	default
parakeet-tdt-ctc-110m	English	220 MB	★
parakeet-tdt-0.6b-v2	English	465 MB
parakeet-tdt-0.6b-v3	25 languages	500 MB

The default is the small one. Set model in the config file if you want a different one. Numbers behind that choice are in docs/benchmarks.md.

They live in the shared FluidAudio cache, at ~/Library/Application Support/FluidAudio/Models. Anything else built on FluidAudio reads the same files, so the download only happens once per machine.

Speed

About 60 ms from key-up to text for a short clip, and roughly twice as quick as Handy on the same model. docs/benchmarks.md has the tables and the commands to reproduce them.

The binary is 9.1 MB with two dependencies. Sitting there waiting for the key it reports 0.0% CPU. Meeting detection is the one thing that ever runs on its own, and that is a single device read a second.

Uninstall

yap install --uninstall
brew uninstall --zap --cask yap

The login item is yap's rather than the cask's, so the first line is what takes it away. Skip it and brew uninstall on its own leaves launchd trying to start a binary that is no longer there. --zap also clears the config file and the logs. Your recordings are never touched, and neither are the models, which are shared with anything else built on FluidAudio.

Requirements

macOS 15 or later on Apple Silicon. Transcription needs the Neural Engine, and recording system audio needs Core Audio process taps.

License

MIT. See LICENSE.