Offline, privacy-first voice typing. Speak into your mic, text appears in your focused app. Or pipe a WAV file and get text on stdout.

Build: Rust, C compiler, cmake, pkg-config, libclang, ALSA headers — sudo apt install build-essential cmake pkg-config libclang-dev libasound2-dev

Run (mic mode): sudo apt install wl-clipboard wtype ydotool — GNOME 45+ / KDE 6.1+ work without wtype/ydotool; pipe mode has no runtime deps

Early stage — free-time side project. Primary target: Ubuntu + GNOME + Wayland. Other distros and compositors are welcome, but I can only test on this setup. If something doesn't work, open an issue. See CONTRIBUTING.md for how to help.

cargo install voicsh

# First run — detects your desktop and picks the best model:
voicsh init

# Mic mode:
voicsh                          # continuous mic → text into focused app
voicsh --once                   # single utterance → exit

# Pipe mode (no mic/runtime deps needed):
cat file.wav | voicsh

voicsh help                     # all commands and options

Mic/WAV → VAD → Chunker → Whisper → Post-processor → Text injection
                                                        ↓
                                                portal / wtype / ydotool

1. Audio captured via cpal (mic) or hound (WAV file)
2. Voice activity detection splits speech into chunks
3. whisper-rs transcribes each chunk locally
4. Text injected via xdg-desktop-portal (GNOME/KDE), wtype, or ydotool

Pipe mode (cat file.wav | voicsh) skips injection and writes to stdout.

# Ubuntu/Debian:
sudo apt install build-essential cmake pkg-config libclang-dev libasound2-dev
cargo install voicsh
voicsh init                    # benchmark hardware, pick model, download
voicsh install-gnome-extension # GNOME Shell panel indicator (optional)

Other distros, GPU acceleration, and pipe-only builds: see INSTALL.md.

voicsh injects transcribed text into your focused app. voicsh init auto-detects the best backend:

Override at runtime: voicsh --injection-backend wtype Override via env: VOICSH_BACKEND=portal voicsh Override in config: [injection] section — run voicsh config dump to see all options.

wl-clipboard (wl-copy) is required for clipboard-based injection.

Note: wtype/ydotool inject via clipboard paste — this overwrites your clipboard. Portal types directly without touching the clipboard.

Voice commands trigger only when spoken as standalone utterances — pause, say the command, pause. Text that merely contains a command word passes through unchanged:

[pause] "period" [pause]          → .
[pause] "new line" [pause]        → (line break)
"the period of history"           → "the period of history"
"press enter to continue"        → "press enter to continue"
[pause] "all caps" [pause] "wow" [pause] "end caps" → "WOW"

Built-in commands are available for English, German, Spanish, French, Portuguese, Italian, Dutch, Polish, Russian, Japanese, Chinese, and Korean — see post_processor.rs for the full list. Discover all commands for a language:

voicsh config list --language=en     # English voice commands
voicsh config list --language=ko     # Korean voice commands
voicsh config list --language=en,de  # multiple languages

Add custom commands in [voice_commands.commands] in config — they take precedence over built-ins. To disable voice commands entirely: voice_commands.enabled = false.

voicsh config dump              # commented template with all options and defaults
voicsh config list              # current active configuration
voicsh config list stt          # just the [stt] section
voicsh config get stt.model     # single value
voicsh config set stt.model small.en

Config file: ~/.config/voicsh/config.toml. Environment overrides: VOICSH_MODEL, VOICSH_LANGUAGE, VOICSH_BACKEND.

GNOME extension — panel indicator with recording state, model info, and Super+Alt+V toggle:

voicsh install-gnome-extension

Shell completions: voicsh completions bash|zsh|fish — run voicsh completions --help for install paths.

MIT

• whisper.cpp - Whisper inference engine
• whisper-rs - Rust bindings
• cpal - Cross-platform audio
• Inspired by nerd-dictation, voxd, and BlahST