Looree

Offline push-to-talk dictation for Windows, powered by NVIDIA Parakeet.

Hold Ctrl+Win, speak, and release. Looree transcribes on your own machine and pastes the text into whatever app has focus. Nothing leaves the computer.

Requirements

• Windows 11
• Python 3.11 or newer
• Optional: an NVIDIA GPU with a current driver

Install

py -m pip install looree

The install takes about 2 GB of disk, most of it the NVIDIA CUDA libraries that GPU support needs.

To work on the code instead, install it from a clone:

git clone https://github.com/digitalblack210/LOOREE.git
cd LOOREE
py -m venv .venv
.venv\Scripts\pip install -e .

Run

pyw -m looree

From a clone, run .venv\Scripts\pythonw -m looree instead.

Looree starts without a console window and adds an icon to the system tray. To have it start when you sign in, right-click the tray icon and tick Start with Windows.

The install also creates a looree.exe command, which does the same thing, but Windows Smart App Control may block it because it is not signed.

The first run downloads the Parakeet TDT 0.6B v3 model (int8 ONNX, about 670 MB) from Hugging Face into the Hugging Face cache. After that Looree runs fully offline.

To watch the log in a console instead, run py -m looree.

Use

• Hold Ctrl+Win and speak. A glow around the edges of the screen shows that Looree is listening and follows the level of your voice.
• Release to transcribe. The glow changes colour while the text is produced, then the text is pasted at the cursor and the glow fades.
• Holds shorter than 250 ms are ignored. Pressing any other key during the hold cancels it, so shortcuts such as Ctrl+Win+D keep working.

Your clipboard is put back after the paste, and transcripts are kept out of the Windows clipboard history.

The tray menu has Open config, Open log folder, Reload config and model, Start with Windows and Quit.

Configuration

Settings live in %APPDATA%\Looree\config.toml, created on first run. Open it from the tray menu, edit, then choose Reload config and model.

The int8 model is built for CPUs and runs fastest there. The log states which provider is active each time the model loads.

Cleanup

Looree tidies each transcript before pasting it. It drops "um", "uh" and "er", and collapses stuttered repeats such as "I I". This needs no extra model and is on by default; set remove_fillers = false under [cleanup] to turn it off.

Words the model does not know

The speech model cannot be given a word list, so Looree fixes names after the fact. Under [vocabulary.hears] in config.toml, put the word you want on the left and the ways the model tends to write it on the right:

[vocabulary.hears]
Acme = ["ack me", "ak me", "akmee"]

Matching is on whole words, in any capitalisation and spacing, so "Ack me's wings" becomes "Acme's wings" and "back me up" is left alone. Copy the wrong spellings from what Looree actually pasted. The bundled config has only a commented example. After editing, choose "Reload config and model" from the tray menu.

Privacy

• To see the hotkey from any app, Looree installs a global keyboard hook. It tracks only the hotkey's own keys; any other key just cancels a hold in progress. No keystrokes are recorded, logged or sent anywhere.
• Audio is processed in memory and never written to disk.
• Transcripts are never logged. The log (%LOCALAPPDATA%\Looree\looree.log) records only timings and character counts.
• With keep_open = true the microphone stream stays open while Looree runs, and Windows shows its microphone indicator. Audio outside a hotkey hold is discarded.

Accessibility

The glow is designed not to flash. Its pulse is slow, its reaction to your voice is smoothed so it follows phrases and not syllables, and it stays lit for at least half a second so tapping the hotkey cannot make it flicker. The test suite simulates the glow and checks it against the WCAG 2.3.1 limit of three flashes per second. This is an automated check of the app's own animation maths, not a medical certification.

Set animate = false under [overlay] for a steady glow with no movement, or enabled = false to turn it off.

Limitations

• Looree cannot see the hotkey while a window running as administrator has focus, unless Looree is started as administrator too.
• Text is inserted with Ctrl+V, so it only works where pasting does.

Transcribe a file

From a clone:

.venv\Scripts\python scripts\transcribe_wav.py path\to\audio.wav

Add --device cpu or --device cuda to choose a provider. Without it the script uses the GPU when one works and the CPU otherwise.

Uninstall

1. Untick Start with Windows in the tray menu, then choose Quit.
2. Run py -m pip uninstall looree, or delete the cloned folder if you installed from a clone.
3. Delete %APPDATA%\Looree (settings) and %LOCALAPPDATA%\Looree (logs).
4. To reclaim the model's disk space, delete its folder from the Hugging Face cache, %USERPROFILE%\.cache\huggingface\hub by default.

Development

.venv\Scripts\pip install -e .[dev]
.venv\Scripts\python -m pytest
.venv\Scripts\python -m ruff check
.venv\Scripts\python -m ruff format --check

The tests run without a microphone, a GPU or the model. Bug reports and pull requests are welcome on the issue tracker.

To release, set version in pyproject.toml, commit, then push a tag of the same version, such as v0.1.0. The release workflow runs the tests, builds the package and publishes it to PyPI.

Credits

Speech recognition uses NVIDIA's Parakeet TDT 0.6B v3 model, released under CC BY 4.0, through the onnx-asr library. The model is downloaded on first run and is not part of this repository.

Looree is an independent project. It is not affiliated with or endorsed by NVIDIA.

License

MIT. See LICENSE.