A lightweight macOS app for voice transcription that stays out of the way: a quote mark in the menu bar, a translucent panel while recording, and finished text wherever the cursor sits. One shortcut (⌥Space by default, configurable) starts and ends a recording; a soft tone marks the moment the microphone is live, and the transcript lands on the clipboard and is pasted into the focused text field. Esc cancels. Three engines to choose from: Whisper through the OpenAI API, Whisper on-device with a local MLX model, or Apple's own on-device speech models.
The menu bar mark shows the current state: ” idle, a red dot recording, … transcribing. While transcribing, the panel shows the waveform of what you just said with a reading head sweeping across it and an ETA learned from your own history — no spinners. Unlike system dictation, nothing is gone after the paste: every transcript is kept in a searchable History window, with lifetime totals for words, audio time, and API cost.
With Homebrew:
brew install --cask lmc4s/tap/verseOr grab Verse-x.y.z-arm64.dmg (Apple Silicon) from the Releases page.
The app is not signed or notarized, so on first launch right-click the app and choose Open, or allow it under System Settings > Privacy & Security. On first use, macOS will also ask for Microphone access, and for Accessibility if auto-paste is on.
OpenAI API — requires an API key and internet. Files are sent to OpenAI's servers.
Local MLX — runs on your Mac, offline, no API key needed. Apple Silicon only. The app manages its own Python environment and pulls models from Hugging Face on first use.
Default model: mlx-community/whisper-large-v3-turbo. Any compatible model from mlx-community works — swap it in Settings.
Apple Speech — macOS's own on-device speech models (macOS 26 or later). Offline, no key, no model download beyond what the system fetches once. Fast and accurate for clear speech in supported locales; Whisper remains the stronger choice for noisy audio and less common languages.
- macOS (Apple Silicon required for the local engine)
- Node.js 18+
- Python 3, Homebrew or system (local engine only)
npm install
npm startnpm run distOpenAI: open Settings from the menu bar icon, paste your API key, select OpenAI as the engine.
Local MLX: open Settings, select Local MLX, click Install. The first transcription also downloads the model weights (1–3 GB depending on the model).
Auto-paste: the first time Verse pastes into another app, macOS asks you to allow it under System Settings → Privacy & Security → Accessibility. Until then it falls back to clipboard-only.
Settings and transcript history are stored in Electron's user data directory (~/Library/Application Support/Verse).
macOS has built-in dictation (the 🎤 / F5 key) that types at the cursor and runs on-device on Apple Silicon. For a quick sentence of clear English it is fast and accurate. Verse covers different ground:
- Whisper holds up on accents, background noise, technical vocabulary, and about 100 languages (Apple's dictation supports roughly 30 locales).
- A choice of engine — OpenAI's API or a fully offline local model — and the model itself is swappable.
- Every transcript is kept: searchable history with lifetime word, duration, and cost totals, one click to copy again.
- A visible recording panel with cancel, live transcription progress with an ETA, a shortcut of your choosing, and the same behavior in every app.
Verse has no telemetry or accounts; the DMG is a packaged build of this repository. Audio is recorded only while a recording is active. The Local MLX engine runs entirely on-device; the OpenAI engine sends audio to OpenAI for transcription. Settings and history are plain JSON files in the user data folder.