yap

Local, private voice dictation for macOS (Apple Silicon). Hold a key, speak, and your words are transcribed on-device and typed wherever your cursor is. No cloud, no account.

Transcription runs entirely on-device via parakeet-mlx (NVIDIA Parakeet TDT 0.6B v3) on Apple's MLX runtime.

Requirements

• Apple Silicon Mac, macOS 14+
• uv

Setup

git clone https://github.com/cluttered-cabinet/yap.git
cd yap
uv sync

Grant your terminal app (Terminal, iTerm, Ghostty, …) these permissions in System Settings → Privacy & Security:

• Microphone — capture audio
• Input Monitoring — detect the dictation key
• Accessibility — type at the cursor

(yap runs from your terminal, so the permissions attach to the terminal, not to yap itself. The first run also prompts you.)

Usage

uv run yap

First run downloads the model (~1.2 GB). A menu-bar item shows the state: yap ⏳ loading · yap ready · yap 🔴 recording · yap ✍ transcribing. (If you see yap ⚠️, Accessibility isn't granted yet — use the menu's "Open Accessibility Settings…", enable your terminal, and restart.)

Two ways to dictate with Right Option (⌥):

• Hold-to-talk — hold ⌥, speak, release. Best for quick phrases.
• Double-tap toggle — double-tap ⌥ to start hands-free recording, double-tap again to stop. Best for long dictation.

Quit from the menu-bar dropdown (or Ctrl+C in the terminal).

Styles (vocal theming)

The menu-bar Style submenu transforms the transcript before it's typed; the choice persists across runs (~/.config/yap/config.json):

• plain — verbatim (default)
• lowercase — all lowercase
• uppercase — ALL CAPS
• clean — LLM-powered cleanup: removes filler words (um, uh, like…), fixes punctuation and capitalization, drops false starts. Uses a small on-device model via mlx-lm (~1 GB download on first use). No cloud, no account.

Add your own in yap/styles.py: any str -> str function in STYLES shows up in the menu automatically.

How it works

hold/2-tap ⌥  →  mic capture  →  log-mel  →  parakeet-mlx  →  keystrokes
(listener      (sounddevice,     (in-RAM,    (on-device      (typed at cursor,
 thread)        16 kHz mono)      no temp     MLX/Metal,       clipboard never
                                  files)      engine thread)   touched)

Three threads: the menu bar (AppKit) owns the main thread, the engine builds and runs the model on its own thread (MLX streams are thread-local), and the keyboard listener runs in the background.

Development

uv run pytest

License

MIT