Yikizi/kuule-kratt-v16b

Kuule Kratt v16b

kuule-kratt-v16b is an experimental Estonian wake-word model for the phrase “Kuule Kratt”. It was developed as part of a bachelor's thesis project on local Estonian voice control for Home Assistant. The model is exported as a TensorFlow Lite artifact with an ESPHome micro_wake_word manifest.

Repositories

• Model development, evaluation tooling, and experimental CLI: https://github.com/Yikizi/kratt-lab
• Home Assistant add-ons, ESPHome examples, and user-facing setup: https://github.com/Yikizi/kratt

Files

• kuule_kratt_v16b.tflite — TensorFlow Lite model.
• kuule_kratt_v16b.json — ESPHome micro_wake_word manifest.
• CHECKSUMS.txt — SHA-256 checksums.

Evaluation summary

The main reported operating point uses threshold 0.97 with the project’s streaming replay protocol.

Metric	Count / exposure	Point estimate	95% CI
Held-out real-speaker recall	139 / 145	95.9%	[91.3, 98.1]
DiPCo meeting-audio FA/h	2 FA / 3.32 h	0.60	[0.07, 2.18]
Home-background FA/h	1 FA / 1.17 h	0.86	[0.02, 4.76]
Common Voice ET FA/h	126 FA / 3.82 h	32.98	[27.48, 39.27]
LibriSpeech FA/h	37 FA / 5.62 h	6.58	[4.64, 9.07]

These measurements place v16b in the useful prototype range for the Kratt project: high recall on the held-out real-speaker control set and low false-activation point estimates on the two recorded background conditions. Dense speech replay and phrase-similarity checks define the current validation boundary.

Validation scope

The model was developed for a low-resource Estonian wake-word prototype. Before using it in a live setting, evaluate it with the intended speaker, microphone, room acoustics, background audio, and phrase variants. Phrase-level checks are especially important for partial phrases, reversed word order, single-word triggers, and phonetically similar continuations.

Data and privacy

This release contains model artifacts, an ESPHome manifest, checksums, and documentation. Training and evaluation audio are retained privately under the project’s consent and privacy constraints.

License

Kratt-owned model artifacts and this model card are released under the Apache License 2.0. Third-party tools and corpora used during development remain under their own licenses and terms.