Kratt Lab is the experimental tooling repository for the Kratt thesis/project. It contains the reusable CLI workflow, Android field-testing apps, and selected wake-word artifacts used to iterate on Estonian local wake-word detection.
For the Home Assistant add-on repository and ESPHome user-facing integration, see https://github.com/Yikizi/kratt.
Status: alpha research tooling. The code and bundled model artifacts are for experimentation and reproducible development, not production-proven voice control.
| Component | Path | Notes |
|---|---|---|
| CLI gateway | cli/kratt |
Small command dispatcher. |
| New wake-word workflow | cli/kratt new-wake-word |
Guided recording, manifest, training, and smoke-benchmark flow for a new phrase. |
| Android false-trigger logger | android/ |
Runs bundled wake-word models and logs activation events on Android. |
| Android dev-voice app | android-dev-voice/ |
Experimental Android-to-Termux voice dispatch; large Sherpa assets are not bundled. |
| Kratt wake-word model archive | wake-word/models/ |
Public TFLite/ONNX artifacts and lineage notes; most versions are historical/diagnostic. |
| Public Korvo-2 firmware | firmware/esphome/ |
Sanitized ESPHome firmware artifacts rebuilt with placeholder credentials. |
./cli/kratt help new-wake-word
./cli/kratt new-wake-word --check
./cli/kratt new-wake-word --list-devices
./cli/kratt new-wake-word --phrase "Hei Toomas" --count 10 --no-stt --no-ttsThe command creates output/new-wake-word/<slug>/ with recordings, prompts,
manifest files, confusable phrase prompts, and follow-up training/benchmarking
commands. Useful models still require real speech/background negatives and a
separate benchmark set.
See docs/new-wake-word/quickstart.md.
The wake-word/models/ directory contains the public Kratt model archive. The
firmware/esphome/ directory contains only sanitized public firmware builds;
private ESPHome binaries can embed Wi-Fi credentials and API keys and must not be
published.
android/README.mddescribes the false-trigger logger.android-dev-voice/README.mddescribes the experimental dev-voice app.docs/android/README.mdsummarizes public-export packaging caveats.
No raw participant audio, user-test WAV files, private Home Assistant logs,
Wi-Fi credentials, or local training datasets are published here. Generated
recordings and training outputs are written under output/ and ignored by git.
Kratt-owned code, configuration, documentation, and project-owned wake-word
model artifacts are published under Apache License 2.0 unless otherwise noted.
Third-party components and downloaded model files remain under their upstream
licenses. See LICENSE and NOTICE.