speak2type

speak2type is a speech-to-text input method engine for Linux desktops, built around IBus integration with a global push-to-talk workflow. It supports both native IBus text input and non-IBus app workflows.

❗ This is less than version 0, use at your own risk!

What it does

• Push-to-talk dictation (default: Alt+Space)

• Two output modes with the same hotkey:

  • IBus/native apps: transcription is committed in place at the current cursor/focus
  • Non-IBus apps (for example VS Code/Electron contexts): transcription is copied to clipboard

• Parakeet backend (local ONNX) as the currently working backend

• Settings app (GTK4 + libadwaita) to manage backend dependencies and models

• Privacy safeguard (IBus-aware fields): recording is disabled for PASSWORD/PIN input purposes

• Global hotkey support via desktop portal listener

• Chrome (Native Ibus) https://youtu.be/QUHK7w-qBb0

• VS Code (Electron, non-Ibus) https://youtu.be/O3MZFxC4ipE

Platform requirements

Minimum expected environment:

• Linux desktop using IBus 1.5+
• Python 3.11+
• GStreamer 1.0
• GTK4 + libadwaita (for preferences UI)

Tested target distro in project docs/scripts:

• Ubuntu 25.10+

Quick start (development install)

The repo includes a development installer that sets up system deps, Python env, IBus component files, and GSettings schema.

./scripts/dev-install.sh

Then ensure your shell exports IBUS_COMPONENT_PATH (if not already set):

export IBUS_COMPONENT_PATH="$HOME/.local/share/ibus/component:${IBUS_COMPONENT_PATH:-/usr/share/ibus/component}"

Restart your shell and IBus:

ibus restart

Finally add Speech To Text as an input source in GNOME Settings.

Debian/Ubuntu package install (APT / .deb)

If you prefer a system package instead of the development installer, the repository includes Debian packaging files under debian/.

Build and install locally:

sudo apt-get update
sudo apt-get install -y debhelper dh-python pybuild-plugin-pyproject python3-all python3-build python3-pytest python3-setuptools python3-wheel
dpkg-buildpackage -us -uc -b
sudo apt-get install -y ../speak2type_0.1.0-1_all.deb
ibus restart

GNOME Settings

For the best push-to-talk experience, adjust these settings in GNOME Settings → Accessibility → Typing:

• Repeat Keys: Off — prevents the PTT key from flooding spaces while held
• Sticky Keys: On — allows modifier+key combos to be registered cleanly by the portal

Usage

1. Select "Speech To Text" as your input source
2. Focus any text field
3. Hold Ctrl+Space and speak
4. Release to transcribe and commit text After install, add Speech To Text as an input source in GNOME Settings.

For detailed packaging steps, see debian/README.build.md.

Usage

1. Select Speech To Text as your active input source.
2. Focus the target app/text field.
3. Hold the push-to-talk hotkey (default Alt+Space).
4. Speak.
5. Release the hotkey to transcribe.
6. Result handling:
   • In IBus-native input contexts: text is inserted at cursor/focus.
   • In non-IBus contexts (e.g. VS Code/Electron): text is copied to clipboard.

Backend status

Working

• Parakeet: fast local ONNX inference, model download supported in preferences.

Experimental / not considered working yet

• Whisper.cpp
• Vosk (gst-vosk)
• HTTP backend (remote/self-hosted API path)

A reference FastAPI server for the HTTP path is provided in server/main.py:

uvicorn server.main:app --port 8000

Endpoints:

• POST /transcribe (generic)
• POST /v1/audio/transcriptions (OpenAI-compatible)

OpenAPI schema: server/openapi.yaml

Development

Install editable package with dev extras:

pip install -e ".[dev]"

Useful commands:

pytest
ruff check src tests
mypy src

Run engine manually (outside IBus integration workflows):

python -m speak2type

Repository layout

• src/speak2type/ — current engine, backends, UI, and model management
• server/ — reference HTTP transcription service + OpenAPI spec
• tests/ — unit/integration tests
• scripts/dev-install.sh — Ubuntu-focused setup helper
• src/upstream/ — upstream reference code retained for comparison/migration

Known limitations

• Alpha quality; crashes/hangs can still occur in edge cases
• Mic symbol may remain stuck as active in task bar (likely GStreamer pipeline stuck in PLAYING)
• Some backend dependencies are large and may require manual system packages
• Vosk path depends on gst-vosk, which may not be packaged on all distros

License

GPL-3.0-or-later. See COPYING.

Credits

This project is based on ibus-speech-to-text from Fedora and continues to evolve as a refactor/new implementation track.