Voice Type

🎙️ Hold-to-talk voice typing for Linux & macOS — hold a key, speak, and the text appears live at your cursor as you talk. Real-time streaming speech-to-text in a tiny Rust tray daemon.

No dictation windows, no clicking — Voice Type turns any focused input box (browser, chat, editor, terminal) into a voice input box. Works on Wayland (sway, Hyprland, GNOME, KDE), X11, and macOS.

• 🗣️ Real-time voice typing — streaming ASR partials are typed live while you speak; no waiting for release
• ⌨️ Push-to-talk hotkey — hold Alt / Option to record, release to finish
• 🔁 Self-correcting — when the recognizer rewrites earlier words, the divergence is backspaced and retyped
• ☁️ Cloud ASR by default — 豆包流式语音识别 2.0 (Seed-ASR) via VolcEngine, two-pass recognition, punctuation & ITN; accurate on Chinese/English mixed speech without any LLM post-processing
• 🏠 Offline fallback — local faster-whisper server when the cloud is unavailable
• 🎨 State-aware tray icon — procedurally rendered, independent of the icon theme
• 🖥️ Cross-platform, single binary — Linux (Wayland/X11) & macOS, no Electron

Paste agent-install.md into your coding agent (pi, Claude Code, Cursor, …) on the target machine — it handles prerequisites, build, API key, autostart and verification:

Install and set up Voice Type (https://github.com/peter209393/voice-type) …
→ full prompt: agent-install.md in the repo root

sudo usermod -a -G input $USER && re-login   # Linux: hotkey access
sudo pacman -S wtype                         # or: sudo apt install wtype

git clone https://github.com/peter209393/voice-type.git
cd voice-type && cargo build --release
export VT_VOLC_API_KEY="your-key"            # or see Configuration
./target/release/voice-type

Get a VolcEngine key at https://console.volcengine.com/speech/new/setting/apikeys?projectName=default, or go fully offline with the faster-whisper server.

Get 豆包流式语音识别模型 2.0 — it is markedly better at English and Chinese/English code-switching. Two ways: (a) enable it for the key's project in the speech console, or (b) subscribe to a 火山方舟 Agent/Coding Plan, turn on 超额后付费 for the speech models, and set the plan's API key as VT_ARK_API_KEY (or VT_ARK_PLAN_API_KEY) — no speech-console key needed at all. Voice Type tries the plan gateway first, then the speech console with 2.0, then 1.0; the startup line [vt] VolcEngine ASR: model 2.0 (seedasr) via ark-plan … confirms which route is active.

Focus any input box, hold Alt and speak — text appears live at the cursor (Chinese/English mixed input works well). Release to finalize. The tray icon shows: idle → recording → transcribing → done/error.

Audio is streamed to the ASR engine in 100 ms packets while you hold the hotkey. Each partial transcript is diffed against what is already on screen; only the delta is typed (with backspaces on rewrites). On release, the final transcript completes the utterance.

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

sway — ~/.config/sway/config:

exec VT_VOLC_API_KEY=your-key ~/.local/bin/voice-type &

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

[Unit]
Description=Voice Type — hold-to-talk voice typing
After=graphical-session.target
[Service]
Environment=VT_VOLC_API_KEY=your-key
ExecStart=%h/.local/bin/voice-type
Restart=on-failure
[Install]
WantedBy=default.target

systemctl --user enable --now voice-type

macOS — add to Login Items (System Settings → General).

MIT