Fast, local-only speech-to-text for macOS. Open source.
EasyVox goes online only when you ask: to download a speech model, and to check easyvox.dev for a newer version.
easyvox.dev · macOS 26 or later, Apple silicon · English
Install
Download the signed, notarized DMG from the latest release, or install EasyVox with Homebrew:
brew install --cask brandonbews/tap/easyvox
EasyVox needs macOS 26 or later on Apple silicon. Or build it from source.
Build from source
You need Xcode or its Command Line Tools, cmake and Rust through rustup.
scripts/fetch-model.sh fetches the default speech model, cargo build and cargo test build and test, and scripts/bundle.sh makes dist/EasyVox.app.
docs/building.md has every step, the end-to-end tests and the self-test.
Settings
EasyVox reads ~/.config/easyvox/config.toml and applies edits when you save; the Settings window writes the same file.
Settings covers the common keys, and the reference lists every one.
Benchmarks
docs/model-bench.md has the measurements behind the model list and the defaults.
Private notes
The maintainer keeps the design decisions, the backlog, release steps and the rules agents follow for this project in a private repo, brandonbews/easyvox-internal, cloned into this checkout as internal/. This repo ignores /internal and /CLAUDE.local.md, and everything builds and tests the same without them. To set up a computer with it:
- Clone this repo, then the private one inside it:
git clone https://github.com/brandonbews/easyvox-internal.git internal.
- Run
sh internal/scripts/install-hooks.sh, and again after moving either checkout. It's safe to rerun, and it refuses to replace a file or a hook it didn't write.
- It links
CLAUDE.local.md here to internal/agent-rules.md, so Claude Code loads those rules in this checkout and in worktrees under .claude/worktrees/.
- It installs a
post-checkout hook that links internal into every new worktree (and CLAUDE.local.md, for a worktree outside this checkout).
License
MIT. Third-party notices: packaging/THIRD_PARTY_NOTICES.html. The speech models keep their own licenses; Settings › About credits each one.
