pomvox is dictation for your Mac. hold one key, say what you mean, and let go — the words show up wherever you were already typing, in email, chat, notes, code, anywhere. no filler words, no fixing punctuation after.
↳ everything happens on your Mac. your voice and your words never leave it — it even works with the wifi off.
macOS 14+ · Apple Silicon (M1+) · notarized
one key, anywhere on your Mac
hit your hotkey and start talking.
the Apple Neural Engine converts speech to text, in real time.
a local model fixes punctuation, casing, and filler words.
polished text lands at your cursor.
There is no network code in the transcription path — your audio and words never leave your Mac.
One-time model download on first setup.
Optional anonymous usage stats: nothing is sent until you explicitly choose on first launch, no field in the payload can carry audio or text, and the code is ~370 lines you can audit in Telemetry.swift.
Because pomvox is open source, you never have to take our word for it.
Pomvox is chapter one of a public investigation into on-device speech inference on Apple Silicon. The same model weights score ~6.3% WER on one runtime and ~2.5% on another — a 3.8-point gap that shouldn’t exist. I’m publishing the hunt for those 3.8 points as I go: the measurements, the dead ends, and what finally moves the number. If low-latency local inference is your kind of rabbit hole, follow along.