Phonon Dictate

Press ⌥Space, talk, press ⌥Space again. The text is on your clipboard.

On-device dictation for the Mac, powered by Phonon-2.

---

Features

• One shortcut. ⌥Space starts and stops recording. Esc cancels. You can change the shortcut in Settings.
• Live ribbon. A pill at the bottom of the screen shows a glowing ribbon, drawn by a Metal shader. Low tones swell it, sharp sounds ripple it, and it gets warmer in color as you get louder. When you stop, it flattens into a line while the model works.
• Clipboard and history. Every transcript goes to the clipboard and to a searchable history window.
• Recordings kept. Each recording is saved as a WAV file, so you can listen to it again.
• Retention periods. Recordings and history are deleted after a period you choose (defaults: 7 days and 30 days).
• Model only in memory while it works. The model loads when you start recording, so it is ready when you stop. It unloads when the transcript is done. Nothing stays in memory between dictations.
• One model. Phonon-2: 164 MB, 5.21 % average word error rate on the Open ASR Leaderboard English sets.
• Number repair. Phonon-2 sometimes writes "like 2.8" as "like2 .8". The app repairs that pattern before it copies the text.

Requirements

• A Mac with Apple silicon and macOS 14 or later
• uv: brew install uv
• Xcode or the Xcode Command Line Tools, to build

Quick start

git clone https://github.com/DevMortimer/phonon-dictate.git
cd phonon-dictate
./scripts/build-app.sh --install    # builds the app and copies it to ~/Applications
open ~/Applications/"Phonon Dictate.app"

First launch

The first launch sets up the speech engine by itself. A setup window shows each step and the live output of the running command:

1. The app makes its own Python environment with uv.
2. It installs the pinned packages from python/requirements.txt (about 1 GB, mostly PyTorch and MLX).
3. It downloads Phonon-2 (164 MB), checks the download, and compiles the GPU shaders.

You can hide the window; setup continues and the menu-bar item shows the progress. When the window says Ready to dictate, press ⌥Space. The first time you dictate, macOS asks for microphone access. If a step fails, the window shows the error, a button to open the setup log, and Try Again.

Usage

Action	How
Start recording	⌥Space
Stop and transcribe	⌥Space
Cancel	Esc
Copy a recent transcript	Menu-bar item → Recent
Search all transcripts	Menu-bar item → History…
Listen to a recording	History → Show Recording

Settings

Open Settings… from the menu-bar item.

Setting	Choices	Default
Shortcut	Any key with ⌃ ⌥ ⇧ or ⌘, or a function key	⌥Space
Keep recordings	1, 7, 30, 90, 365 days, or forever	7 days
Keep history	1, 7, 30, 90, 365 days, or forever	30 days
Launch at login	On or off	Off

Privacy

All recording and transcription happen on your Mac. The app connects to the internet only during the first launch, to download the Python packages from PyPI and the model from Hugging Face. After setup, the worker runs with HF_HUB_OFFLINE=1.

Files

Everything is in ~/Library/Application Support/Phonon Dictate/:

Path	Content
recordings/*.wav	16 kHz mono recordings
history.json	Transcripts
python/	The Python environment
setup.log	Log of the first-launch setup
worker.log	Log of each transcription

The model files are in ~/.cache/fermion/.

How it works

⌥Space ──► Recorder (AVAudioEngine, 16 kHz WAV) ──► overlay waveform
   │
   └─────► worker.py starts and loads Phonon-2 while you speak
⌥Space ──► WAV path sent to the worker ──► text ──► clipboard + history
                                            └──► worker exits, model unloads

• Sources/PhononDictate/: the Swift app. Carbon global hotkey, recorder, overlay panel, menu-bar item, settings, and history.
• python/worker.py: a one-shot worker. It loads Phonon-2 through the fermion-research package, reads one WAV path from stdin, prints one JSON line, and exits.
• Sources/PhononDictate/Ribbon.swift: the ribbon shader. The recorder splits each audio buffer into low, mid, and high bands with an FFT (Accelerate), and the shader turns them into motion and color.

Development

swift build                                   # debug build
python3 -m unittest discover -s python        # worker tests
./scripts/build-app.sh                        # app bundle

Troubleshooting

Problem	Fix
The menu says the shortcut is used by another app	Choose a different shortcut in Settings.
No microphone prompt appears	Run tccutil reset Microphone com.devmortimer.PhononDictate, then dictate again.
"Transcription failed"	Read worker.log.
The setup window shows "Setup stopped"	Click Open Setup Log, fix the cause, then click Try Again.

Uninstall

rm -rf ~/Applications/"Phonon Dictate.app" ~/Library/Application\ Support/Phonon\ Dictate ~/.cache/fermion
defaults delete com.devmortimer.PhononDictate

License

MIT for the code in this repository.

Credits

• Phonon-2 by FermionResearch. Weights: CC-BY-4.0. Based on NVIDIA parakeet-tdt-0.6b-v3.
• fermion-research: Apache 2.0.