Voice Type

A cross-platform voice-to-text application that types what you speak. Hold a hotkey to record, release to transcribe and type.

Features

• Cross-platform: Linux (X11/Wayland) and macOS
• Native UI: State-aware system tray icon + recording popup near cursor
• Local ASR: On-device speech recognition with faster-whisper (no cloud, fully offline)
• Smart correction: Optional post-processing of the transcript with a local LLM (qwen2.5 via Ollama) to fix homophones, punctuation and filler words
• Fast feedback: Raw transcript is typed immediately; the corrected version replaces it in the background once ready
• Global hotkey: Hold Alt/Option to record, release to type

Supported Platforms

Platform	Hotkey	Typing	Tray	Popup
Linux (Wayland)	Alt	wtype, ydotool	ksni	GTK
Linux (X11)	Alt	xdotool, ydotool	ksni	GTK
macOS	Option	enigo	NSStatusItem	NSWindow

Requirements

Linux

• input group membership for hotkey detection:

  sudo usermod -a -G input $USER
  # Log out and back in for changes to take effect

• Typing tool (install one):

  # Wayland (recommended)
  sudo pacman -S wtype      # Arch
  sudo apt install wtype    # Debian/Ubuntu
  
  # X11
  sudo apt install xdotool
  
  # Universal (works on both)
  sudo pacman -S ydotool    # Arch
  yay -S ydotool

macOS

• Accessibility permissions:
  • System Settings → Privacy & Security → Accessibility
  • Add the app to the list

Dependencies (ASR + Correction)

Voice Type runs two local services. Install them once:

1. Ollama + qwen2.5 (transcript correction)

curl -fsSL https://ollama.com/install.sh | sh
ollama pull qwen2.5:3b-instruct   # fast on CPU; use qwen2.5:7b-instruct for quality

The installer usually registers an ollama systemd service that auto-starts. Verify it is listening on port 11434 (curl http://127.0.0.1:11434/api/tags); if not, start it manually with ollama serve (or systemctl start ollama).

Prefer non-thinking instruct models here (e.g. qwen2.5). Thinking models like qwen3 ignore the "output only the corrected text" instruction and emit long chains of thought, which is both slow on CPU and pollutes the output.

2. faster-whisper ASR server

The server runs in its own isolated Python (3.12) via uv, so your system Python is not affected. Install uv, then from the repo root:

uv run --python 3.12 \
    --with fastapi --with "uvicorn[standard]" \
    --with "faster-whisper" --with python-multipart \
    uvicorn server.asr_server:app --host 127.0.0.1 --port 8000

The first run downloads the ASR model (~1.5 GB for medium, cached under ~/.cache/huggingface). On a CPU-only machine the default medium/int8 model is a good speed/accuracy trade-off; see VT_ASR_MODEL below to change it.

The ASR server must be running whenever Voice Type is. Consider wrapping the command above in a systemd unit alongside Voice Type (see Autostart).

Installation

From Source

git clone https://github.com/peter209393/voice-type.git
cd voice-type

# Linux (default, ksni tray)
cargo build --release

# Linux (GTK tray - for XFCE, MATE, etc.)
cargo build --release --features gtk-tray

# macOS
cargo build --release

Usage

1. Start the ASR server (see Dependencies) and ensure Ollama is running.

2. Run the application:

   ./target/release/voice-type

3. Hold Alt (Linux) or Option (macOS) to start recording

4. Speak your text

5. Release the key: the raw transcript is typed immediately, then replaced seconds later by the corrected version

Autostart

Linux (Sway)

Add to ~/.config/sway/config:

exec /path/to/voice-type

Linux (systemd)

Create ~/.config/systemd/user/voice-type.service:

[Unit]
Description=Voice Type
After=graphical-session.target

[Service]
ExecStart=/path/to/voice-type
Restart=on-failure

[Install]
WantedBy=default.target

A companion unit for the ASR server (voice-type-asr.service) is recommended, running the uv run ... uvicorn ... command from Dependencies.

Then:

systemctl --user enable --now voice-type

macOS

Add to Login Items in System Settings → General → Login Items.

Configuration

Environment Variables

Variable	Default	Description
VT_ASR_URL	http://127.0.0.1:8000	faster-whisper ASR server base URL
VT_ASR_MODEL	medium	faster-whisper model (tiny/base/small/medium/large-v3)
VT_ASR_DEVICE	cpu	CTranslate2 device (cpu/cuda)
VT_ASR_COMPUTE_TYPE	int8	CTranslate2 compute type (use float16 on CUDA)
VT_CORRECT	1	Enable (1) / disable (0) LLM correction
VT_CORRECTOR_URL	http://127.0.0.1:11434	Ollama base URL
VT_CORRECTOR_MODEL	qwen2.5:3b-instruct	Correction model (e.g. qwen2.5:7b-instruct for quality). Avoid thinking models like qwen3 — they ignore "output only" and are slow on CPU

Building Features

Feature	Description	Default
gtk-tray	Use GTK/libappindicator for tray (XFCE, MATE)	No

Troubleshooting

Problem	Fix
"No keyboard device found"	Add yourself to the input group and re-login
"No typing tool found"	Install wtype, ydotool, or xdotool
ASR errors / connection refused	Start the faster-whisper server (see Dependencies); check VT_ASR_URL
"Corrector request failed"	Ensure Ollama is running and the model is pulled (ollama pull qwen2.5:3b-instruct); or set VT_CORRECT=0
Correction is too slow	Use a smaller model or VT_CORRECT=0
Text gets replaced wrongly	An IME can break backspace-based replace; disable correction (VT_CORRECT=0) or the IME while typing
No tray icon	Ensure your bar supports StatusNotifierItem (waybar, etc.)
macOS: Typing doesn't work	Grant Accessibility permissions in System Settings

License

MIT