Ryan Vogel a149464ac8 Android: filter keyboard candidates without rewriting typed words
Replace unconditional HXTM output with lookup-time flags for prefix and next-word suggestions, deterministic fixes, whole-word spatial readings, and both halves of missed spaces. Still use blocked words for probabilities, and keep a fully typed spelling (including out-of-model words) through space or a distinct literal chip. The Inter Keyboard switch enables profanity only; preference-aware caches refresh even in the active editor.

Verified 8 filter tests and 5 existing tests in each of debug and release, including every curated entry across the four layouts. Reproduced offensive suggestions and space replacements on the Studio API 35 IME before the change; after it, confirmed safe near-miss results, intentional literals, live toggle refresh, and light/dark settings. The 4,000-sentence phase-4 replay changed English WER from 7.4045% to 7.4117% and left Spanish at 10.3301%.
2026-11-20 13:02:31 -05:00
2026-11-02 14:11:35 -05:00
2026-11-10 17:51:36 -05:00

Hex

Hex is local voice dictation for Mac, Windows and Linux. Hold a key, speak, and it pastes.

How it works

  • hold the shortcut, speak, then let go
  • Hex transcribes on your device and pastes into whatever app you're in
  • the first run downloads a speech model; after that nothing leaves your machine
  • experimental features live in Labs, at the bottom of the sidebar

What it does

  • hold the shortcut to dictate, or set Shortcut mode to Tap to start and stop with a press
  • the text lands at the cursor in whatever app you're in
  • Option-Shift-V pastes the last dictation again (Ctrl-Alt-V on Windows and Linux)
  • History keeps your dictations, searchable, for 7 days by default
  • Output runs every dictation through steps you choose (word replacements, built-in cleanups, your own scripts, an optional OpenCode rewrite), with different steps per app or website
  • Vocabulary is one list of names Hex should know; Whisper and Qwen3-ASR read it while transcribing, Parakeet models read it with Labs › Vocabulary help on, and the OpenCode rewrite uses it
  • Annotate records the screen while you draw and talk, on macOS, Linux and Windows (GNOME Wayland needs Labs › Annotate on GNOME)
  • an authenticated local API and the TypeScript SDK transcribe with the same model, for your own tools
  • Meetings is a macOS lab that records a call and writes the transcript

Output and History:

Hex output: the steps every dictation runs, a Try box, and different steps per app Hex history: past dictations, a search field, and one entry's raw transcript, app, processing and model

Speech models

Hex downloads one model the first time it runs and transcribes with it on your machine from then on. On an English system that's Parakeet Unified English: it writes while you talk, so the text is ready when you let go. Other languages start with the model recommended for them. That model also serves meetings and hex transcribe.

Settings' picker offers more, among others:

  • Parakeet v3: 25 European languages, and it hears which one you speak
  • Cohere Transcribe: the fewest wrong words, in 14 languages; a big download
  • Whisper large-v3-turbo: almost any language, and it reads your vocabulary
  • Qwen3-ASR 0.6B: Chinese and 29 more, and it reads your vocabulary
  • SenseVoice Small: Chinese, Japanese, Korean and English; a small download
  • Nemotron 3.5 ASR: writes while you talk, in 27 languages
  • Granite Speech 5.0: IBM's model, seven languages, with punctuation
  • one-language models: ivrit.ai Whisper for Hebrew, GigaAM v3 for Russian, NB-Whisper for Norwegian, KB-Whisper for Swedish, QwenCleo-ASR for Egyptian Arabic
  • more English models: Parakeet v2 and Underdog Bark 3.0

Add a GGUF model of your own with "Add model file…" in the picker; see your own speech models.

Downloads

Stable is the current release. Dev is the newest build of dev and updates itself from the dev channel.

Platform Stable Dev
macOS (Apple silicon) DMG DMG
Windows x64 Installer Installer
Windows ARM64 Installer Installer
Linux x86_64 Executable Executable
Linux x86_64 (Debian, Ubuntu) .deb .deb
Linux x86_64 (Fedora, openSUSE) .rpm .rpm
Linux ARM64 Executable Executable
Linux ARM64 (Debian, Ubuntu) .deb .deb
Linux ARM64 (Fedora, openSUSE) .rpm .rpm

Or install with:

  • Homebrew: brew install --cask anomalyco/tap/hex
  • AUR: yay -S hex-bin
  • install script (Linux): curl -fsSL https://anoma.ly/hex/install-linux.sh | sh
  • Nix (Linux): nix profile add github:anomalyco/hex

Command line

hex transcribe meeting.m4a prints a recording's transcript, whether or not the app is running. It reads WAV, AIFF, CAF, FLAC, MP3, Ogg, WebM, and MP4/M4A/MOV, including a video's soundtrack. The same hex controls the running app (hex status, hex dictation, hex settings); see docs/cli.md.

On macOS the CLI is inside the app. Homebrew links it for you; otherwise link it somewhere on your PATH that you own:

ln -s /Applications/Hex.app/Contents/MacOS/hex ~/.local/bin/hex

Linux and Windows installs put hex on your PATH.

More docs are in docs/.

Contributing

We don't take pull requests. Open an issue describing the bug or idea, and our agents usually turn it around fast.

MIT licensed.

Languages
Rust 83.5%
Swift 7.8%
JavaScript 2.1%
HTML 1.4%
TypeScript 1.2%
Other 3.7%