TypeFlow

A local, private voice-to-text app for macOS. Press ⌘\ anywhere, speak, release — your transcription lands on the clipboard, ready to paste.

Runs entirely offline using Whisper via faster-whisper. No accounts, no cloud, no telemetry.

Features

• Global hotkey — ⌘\ to start/stop recording from any app
• Local Whisper — choose from small (240 MB, fast) up to large-v3-turbo (1.5 GB, ChatGPT-grade)
• English + Indonesian — selectable from the menu bar (Whisper supports 99 languages — only EN/ID exposed in the UI)
• Floating overlay — ChatGPT-voice-mode-style pill at the bottom of your screen, with live audio levels
• Main window — animated orb showing current state + last transcript
• Menu bar item — quick access to language, model, microphone
• Microphone picker + test mode — works with any input device on your system
• Fully local — model weights cached at ~/.cache/huggingface, used offline forever after

Requirements

• macOS 13 (Ventura) or later. Tested on macOS 14 / 15.
• Python 3.11+ (only needed to build from source — the bundled .dmg ships its own Python)
• ~2 GB of disk for the largest model

Quick start (build & run from source)

git clone https://github.com/sanwidee/typeflow.git
cd typeflow
./run.sh

run.sh creates a virtualenv, installs dependencies, and launches the app. The first transcription downloads the Whisper model (one-time, ~240 MB for small).

Build a .app for your menu bar

./build_app.sh

This produces ~/Applications/TypeFlow.app — a thin launcher that points at your local venv. Double-click to run. macOS will prompt for:

• Microphone (for recording)
• Accessibility + Input Monitoring (for the global hotkey)

Grant all three, then quit and relaunch the .app.

Build a distributable .dmg

./scripts/build_release.sh

This uses PyInstaller to produce a fully self-contained .app that bundles its own Python + dependencies (no Anaconda or system Python needed). Then:

./scripts/make_dmg.sh

…packages that .app into dist/TypeFlow.dmg, ready to upload to GitHub Releases.

Installing an unsigned .dmg

The DMG is not signed by Apple (paying $99/yr for a Developer ID isn't reasonable for a free hobby project). The first time someone downloads it, macOS will refuse to open it. Two ways to bypass:

Option A — right-click → Open (recommended):

1. Drag TypeFlow.app to /Applications
2. Right-click → Open
3. Click Open in the warning dialog
4. After this, double-click works normally

Option B — strip the quarantine flag (for "App is damaged" errors):

xattr -dr com.apple.quarantine "/Applications/TypeFlow.app"

Usage

Action	How
Start/stop recording	⌘\
Change language	Menu bar TF → Language
Change Whisper model	Menu bar TF → Model
Pick microphone	Menu bar TF → Microphone
Test microphone (3s)	Menu bar TF → Test Microphone
Show main window	Menu bar TF → Show Window, or click the Dock icon
Quit	Menu bar TF → Quit
Logs	~/.typeflow.log
Settings file	~/.typeflow.json

Architecture

app.py          — main entry point, rumps menu bar, state machine
main_window.py  — animated NSWindow with orb, status, transcript
overlay.py      — floating NSPanel pill (ChatGPT-voice-mode style)
build_app.sh    — wraps the venv into a clickable .app
assets/icon.svg — source for the .icns app icon
scripts/
  build_icon.sh    — SVG → .icns via qlmanage + sips
  build_release.sh — PyInstaller bundle (self-contained .app)
  make_dmg.sh      — wraps the .app into a .dmg
  install_launcher.sh — internal helper for build_app.sh

License

MIT — see LICENSE.

This project bundles open-source components under MIT, BSD, Apache-2.0, and LGPL-3.0. See NOTICE for full attribution and LGPL compliance details.

Acknowledgements

• OpenAI for Whisper
• SYSTRAN/faster-whisper for the fast CTranslate2 backend
• Inspired by Wispr Flow, Superwhisper, and ChatGPT voice mode