Wisp

Hold a key, talk, let go, and the text is there — cleaned up, punctuated, in whatever app you were already in. Nothing leaves the machine.

Your thoughts outrun your hands. Wisp closes the gap: hold a key anywhere on the Mac, say what you mean, let go, and finished text appears at your cursor, in Mail, in Slack, in the editor you were already in.

Speech is messy. Nobody wants a transcript.

Most dictation hands back exactly what you said, false starts and all, so you spend the time you saved fixing commas. Wisp runs the raw transcript through Apple’s on-device model first. Filler words go, punctuation appears, and saying “new paragraph” produces a paragraph rather than the words. Your phrasing survives: it is a typist, not a ghostwriter.

Select something first, and it rewrites instead

There is no command word and no mode switch. When you press the key, Wisp checks whether the app you are in has text selected. Nothing selected, and you are dictating. Something selected, and your words become an instruction for that text (make this a bullet list, cut it to two sentences) and the selection is replaced.

How it works

1. Hold the key. Right ⌥ by default. Wisp notes which app you are in and what you have selected.
2. Talk. Words appear in the capsule as they are recognised, so you can see it is hearing you.
3. Let go. The cleanup pass runs and the text is pasted. Your clipboard is put back exactly as you left it.

Why on-device

Dictation is the most personal text you produce: half-formed thoughts, names, notes you would never paste into a website. Running the model locally is not a feature, it is the point.

Parakeet handles transcription, Apple’s Foundation Models clean the raw output into something you would send. Both run on the machine, so it works on a plane and it works with the Wi-Fi off.

Nothing leaves the machine

This is not a privacy policy promising deletion: there is no server to delete anything from. No account, no telemetry, and no network calls after the one-time model download.

Stage	Runs on
Capture	AVAudioEngine, 16 kHz mono
Recognition	Parakeet TDT v3, Neural Engine
Cleanup	Apple Intelligence, on-device
Insertion	Pasteboard, restored afterwards
History	JSON in Application Support

Install

Version 0.1.1 needs macOS 26 on Apple silicon. It’s an 8.4 MB installer and it’s free. Download it and open it: Wisp lands in Applications and opens straight into a welcome screen that walks you through the two permissions it needs (microphone, and accessibility so it can see your hotkey and type into other apps), then downloads its speech model once.

Where it is

Daily-driver quality for me, not yet for anyone else.