Transcribe

Dictation and file transcription for macOS 26, on Apple's on-device speech stack. 448,464 bytes. No models, no Python, no ffmpeg, no server.

Measured on an M4, warm: paste fires 58–105 ms after key-up (p50 58 ms). A 4-minute file transcribes in 4,200 ms — RTF 0.0176×.

Use

• ^Space (configurable): dictate; the transcript pastes into the focused app. The capsule pill rides below the notch — a live waveform while you speak, "Transcribing…" while it thinks, a green check when done. Click or Esc cancels.
• Drop a WAV/AIFF/CAF/M4A on the menu-bar mic (or Finder Quick Action): a <file>.md transcript appears beside it.
• Language menu: Auto, or any language this Mac's speech stack offers.
• Nothing persists. A dictation lives in your clipboard; a file transcript lives in the .md beside the audio. The app forgets the rest.

Install

Unzip, drag Transcribe.app to /Applications. Apple Silicon, macOS 26+. Releases: https://github.com/gbrlpzz/transcribe/releases

CLI

The app binary is the CLI — one command:

sudo ln -sf /Applications/Transcribe.app/Contents/MacOS/Transcribe /usr/local/bin/transcribe

transcribe notes.m4a --json
# {"file","text","language","elapsed_ms","md_path"}

Missing language assets install automatically. Exit codes: 0 ok · 2 usage · 3 unreadable file · 4 language not ready · 5 failed.

Build

make app     # app/dist/Transcribe.app
make dist    # release zip
make test    # 32-case battery + dictbench latency harness

Privacy

On-device. No analytics; network only for the manual update check. docs/PRIVACY.md · docs/APPLE-SPEECH-API-NOTES.md · docs/TROUBLESHOOTING.md

License

MIT — see LICENSE. Copyright (c) 2026 Gabriele Pizzi