Rust · whisper.cpp · macOS, Linux, Windows · MIT

Halfvoice

System-wide voice dictation. 100% on your machine.

Hold a key anywhere in your OS and speak. Release, and the transcribed, cleaned-up text is typed into whatever app has keyboard focus.

Install GitHub

right option held · listening

words appear as you speak

Hold a key
Speak
Release

Install

macOS

Apple Silicon, Metal-accelerated.

brew tap aimingeye/tap
brew install halfvoice

Non-Homebrew downloads aren’t notarized. After extracting, run xattr -d com.apple.quarantine ./halfvoice and verify with shasum -a 256 -c SHA256SUMS.

Linux

X11. Optional CUDA and Vulkan builds are published alongside the default one.

Grab a build from the Releases page, then verify it against the published SHA256SUMS:

sha256sum -c --ignore-missing SHA256SUMS

Building from source: see the README.

Windows

Optional CUDA and Vulkan builds are published alongside the default one.

Grab a build from the Releases page, then verify it against the published SHA256SUMS:

sha256sum -c --ignore-missing SHA256SUMS

Run that in Git Bash or WSL. Building from source: see the README.

Quickstart

halfvoice models download base.en   # one-time, ~148 MB
halfvoice setup
halfvoice run

Hold Right Option (macOS/Linux) or Right Ctrl (Windows), speak, release.

Features

Push-to-talk or toggle mode.
Streaming dictation: words appear while you’re still speaking.
Works in any app that has keyboard focus.
Spoken commands: “new line”, “scratch that”, spoken punctuation.
Custom vocabulary biases recognition toward your names and jargon.
Filler-word removal plus a personal replacement dictionary.
Voice notes appended to a daily Markdown file.
halfvoice transcribe memo.wav for audio files.

Privacy

Audio lives only in memory. It never touches disk and it never touches the network. No telemetry, no analytics, no update checks.

Halfvoice runs fine with networking disabled. The only network call it ever makes is the explicit one-time model download you run at setup.