Yo-Voice

Yo-Voice (Ёхо) is local push-to-talk dictation for Linux. Press the hotkey, speak, pause — the transcript is pasted into whatever text field has focus.

It is not a cloud product. Microphone audio and transcripts stay on the computer that runs it.

What it does

• Types what you say into the currently focused text field (browser, editor, chat, terminal).
• Runs on your machine. Speech is recognized locally with faster-whisper (large-v3-turbo). Nothing is uploaded to a speech API.
• Speaks Russian. Punctuation and capitalization are cleaned up after recognition.
• Shows a small cat overlay while listening: Жду (waiting / loading), Можно говорить (about one second after the mic is live), then Слушаю (listening). The waveform moves only while the microphone is actually capturing.
• Lets you pick the microphone by right-clicking the cat.
• Starts this program at login, so the hotkey is ready without opening a window.

What it cannot do

Yo-Voice 1.0.0 does not:

• Run on Windows or macOS
• Run on a Wayland-only session (it needs X11 for the global hotkey and paste)
• Send audio or text to the cloud, or offer an account / sync
• Stream words into the cursor while you are still talking (it pastes after a pause, or when you press the hotkey again)
• Keep a history of recordings or WAV files
• Switch languages (recognition is locked to Russian)
• Replace a full accessibility on-screen keyboard or a system-wide IME

Developed and tested on Linux Mint 22.3, Cinnamon, X11. Other X11 desktops may work; they are not the supported target.

Privacy

• Your speech never leaves this computer. Recognition runs locally.
• Transcripts are not uploaded. They are pasted into the focused app through the clipboard, then the previous clipboard contents are restored.
• Config lives in ~/.config/yo-voice/. Logs live in ~/.cache/yo-voice/.
• The only expected network use is install time (Python packages) and the first launch, which downloads Whisper model weights (~1.6 GB) into a local cache. That download is model files, not your microphone.

Hotkey

Dictation is bound to the key immediately left of 1:

Keyboard layout	Key
Russian (ЙЦУКЕН)	ё
US / UK	` (grave / tilde key, unshifted)

X11 keycode 49. Holding Shift+ё still types Ё — only the unshifted key is grabbed.

If another app already owns that key, the overlay says клавиша ё занята.

Requirements

• Linux with an X11 session (Cinnamon on Linux Mint 22.3 is the reference)
• Python 3.10+
• A working microphone
• NVIDIA GPU strongly recommended (tested on RTX 4070 Ti). CPU fallback exists and is slow.
• Disk: a few hundred MB for the app, plus ~1.6 GB for the Whisper model on first run

Install on Linux

1. System packages

sudo apt update
sudo apt install -y \
  git curl \
  python3 python3-venv python3-pip \
  python3-gi python3-gi-cairo python3-cairo python3-xlib \
  gir1.2-gtk-3.0 libportaudio2

Install your NVIDIA driver from Driver Manager if you want GPU recognition.

2. Download

git clone https://github.com/iammedved/Yo-Voice.git
cd Yo-Voice

3. Install

chmod +x scripts/install.sh
./scripts/install.sh

The installer:

1. Creates .venv (with system GTK / X11 bindings)
2. Installs Python dependencies, including CUDA libraries when available
3. Puts a Ёхо launcher on the Desktop and in the application menu
4. Enables autostart of this program (~/.config/autostart/yo-voice.desktop, 3 second delay)
5. Symlinks yo-voice into ~/.local/bin/

4. First launch

1. If the Desktop icon asks for permission, choose Allow Launching.
2. Click Ёхо, or run yo-voice (add ~/.local/bin to PATH if needed).
3. Press ё / `. The cat appears. The first time, it downloads the model and the overlay may show a loading status for a minute.
4. Speak Russian. When you pause, text is pasted at the caret. Press the hotkey again to stop.

Right-click the cat to choose the microphone (Авто = automatic).

Later launches (every day after that)

You do not need to open the app by hand.

• After you log into Cinnamon, this program starts by itself.
• Press ё / ` to start or stop listening.
• The Desktop / menu icon toggles listening the same way.
• First-run model download happens only once; later starts are local and fast on GPU.

Useful commands:

yo-voice              # toggle listening (starts this program if needed)
yo-voice status       # idle | listening
yo-voice start        # start listening
yo-voice stop         # stop and paste
yo-voice settings     # microphone window
yo-voice quit         # quit this program
yo-voice demo         # overlay demo without dictation

Re-run ./scripts/install.sh after a git pull to refresh the venv and launchers.

How a dictation pass works

1. Hotkey (or the launcher) tells this program to listen.
2. Overlay appears without stealing keyboard focus, so the caret stays in your app.
3. Silero VAD keeps silence out of Whisper; only speech is transcribed.
4. On a pause, the utterance is recognized locally and pasted with Ctrl+V (Ctrl+Shift+V in terminals).
5. Press the hotkey again to flush the last bit and hide the cat.

Troubleshooting

Symptom	What to try
Waves do not move	Unmute the headset mic. Right-click the cat and pick the real device, not a monitor/loopback.
подключите микрофон	No usable capture device. Plug in a mic and pick it in the menu.
клавиша ё занята	Another program grabbed keycode 49. Close it or change that grab.
Icon does nothing	Desktop files on Mint often need Allow Launching. Or run yo-voice in a terminal.
First start is slow	Model download / GPU load. Later starts should not download again.
GPU unused	NVIDIA driver + the nvidia-cublas-cu12 / nvidia-cudnn-cu12 wheels from install.sh. CPU still works, slowly.
Nothing pastes	Click the target field first. The overlay never takes focus on purpose.

Versioning

This tree is 1.0.0. See CHANGELOG.md. Git tags follow semver (v1.0.0).

License

MIT. Third-party models (faster-whisper / Whisper weights, Silero VAD) keep their own licenses.