Kratt is a research prototype for Estonian Home Assistant voice control: a "Kuule Kratt" ESPHome wake-word model plus modular Estonian STT/TTS components for Home Assistant Assist.

Status: v0.1.0-alpha research prototype. The v16c wake-word model is a stable demo/baseline model, not a production-proven detector. Later thesis experiments showed unresolved prefix/confusable phrase-selectivity risks.

Add this repository in Home Assistant:

Settings → Add-ons → Add-on Store → ⋮ → Repositories
https://github.com/Yikizi/kratt

Install whichever add-ons you need:

• Kratt Kiirkirjutaja STT — local Estonian speech-to-text.
• Kratt Neurokõne TTS — Estonian text-to-speech through Neurokõne.

For the wake word, add the public ESPHome model manifest to an ESPHome voice satellite:

micro_wake_word:
  models:
    - model: github://Yikizi/kratt/wake-word/models/kuule-kratt-v16c/kuule_kratt_v16c.json@main
      id: kuule_kratt_model

See docs/user-guide/home-assistant-quickstart.md and home-assistant/README.md.

Wake-word detection runs locally on the ESP32-S3 class device. The Kiirkirjutaja STT add-on processes audio locally after downloading upstream model files. The Neurokõne TTS add-on sends synthesis text to the external TartuNLP API; use Piper or another local TTS if a fully offline stack is required.

No raw participant audio, user-test WAV files, private Home Assistant logs, Wi-Fi credentials, or local training datasets are published in this repository. See PRIVACY.md.

The thesis work behind this repository found that clip-level wake-word accuracy is not enough for real deployment. Models must be evaluated together on:

• streaming false accepts per hour (FAPH),
• real/unseen-speaker recall,
• hard-negative, prefix, and confusable phrase rejection.

The v16c model is published as a baseline/demo artifact. It should not be marketed as a production-ready Estonian wake-word detector. See:

• wake-word/models/kuule-kratt-v16c/NOTES.md
• wake-word/docs/MODEL_LINEAGE.md
• docs/evaluation.md

Kratt-owned code, configuration, documentation, and wake-word model artifacts are published under the Apache License 2.0 unless otherwise noted. Third-party components and downloaded model files remain under their upstream licenses. See LICENSE and NOTICE.