EXPERIMENTAL v0.0.0 · APPLE SILICON macOS · BUILD FROM SOURCE

Hold a key. Speak. Release.

Your words land at the cursor of the app you’re already in. No dictation window, no copy-paste.

How it works

One gesture. One destination. Two routes to get there.

Hold the Talk Key (Right Option) and one Utterance opens. Speak; on release, type-wave commits the Final Transcript — the only text it ever inserts — and places it at your Focused Target, the cursor of the app that already has focus. One hold, one insertion. A silent HUD shows a red waveform while you talk, then green processing dots until the text lands — it never shows the transcript itself.

OpenAI Realtime — the default

Streams your voice to OpenAI's Realtime API over an encrypted WebSocket and returns text as you speak — the fastest route. Needs an OpenAI API key with Realtime access. Your audio streams to OpenAI while you dictate.

Local Whisper — offline

Runs Whisper Large v3 Turbo entirely on your Mac. No API key, no network while you dictate — the audio never leaves your machine. A one-time, credential-free model download and you're set.

type-wave demo: hold the Talk Key, speak, release, and the Final Transcript lands at the cursor of a TextEdit window.

Your voice can stay on your Mac

Switch to local Whisper and type-wave goes fully offline.

Point the Backend Router at local Whisper and type-wave stops touching the network. Your voice is captured, transcribed, and inserted without a byte leaving your Mac — no API key, no cloud, no dictation-time network. The only fetch is a one-time, credential-free download of the pinned model; after that, local mode runs with no network at all. (In the default OpenAI mode, your audio does stream to OpenAI while you dictate — privacy is what you get when you choose local.)

Build it and run it

From source with Nix and Zig on an Apple Silicon Mac. No binary to download — you build it yourself.

Apple Silicon Mac — Intel Macs aren't supported
Nix with flakes (or a matching bare Zig nightly)
macOS permissions on first run: Input Monitoring (observe the Talk Key), Accessibility (insert text), Microphone (capture audio)
OpenAI mode: an OpenAI API key with Realtime access
Local mode: nothing extra — just the one-time, credential-free model download

Try it (foreground dev run)

export OPENAI_API_KEY=sk-...
nix develop --command zig build run

Install for daily use (LaunchAgent)

nix develop --command zig build install-agent
~/.local/bin/type-wave --set-key

Prefer offline? Choose the local backend and you can skip the API key.

HOW IT WAS BUILT

Charted with wayfinder, part of Matt Pocock's AI Skills for Real Engineers. The plan for this site was mapped as decision tickets and resolved one at a time before a line was built.