Files
ARIA-AGENT/android/android/app/src/main/assets/openwakeword/computer.onnx
T
duffyduck 309df9d851 fix(wake-word): Embedding-Output ist rank-4, nicht rank-2 — Trigger funktioniert jetzt
Hauptursache warum kein Wake-Word je triggerte: das Google-Speech-
Embedding-Modell liefert (1,1,1,96), nicht (1,96). Der Cast
`as Array<FloatArray>` warf eine ClassCastException, die vom try/catch
geschluckt wurde — Pipeline lief still ins Leere.

Zusaetzlich:
- WW-Input-Frame-Count wird jetzt aus den Modell-Metadaten gelesen
  (variiert pro Keyword; hey_jarvis=16, computer_v2evtl. anders)
- "Computer" als Wake-Word erweitert (Community-Modell aus
  fwartner/home-assistant-wakewords-collection)

"ARIA" als Wake-Word: gibt's nicht fertig trainiert. Muesste ueber
das openWakeWord Colab-Notebook trainiert werden (~1h auf gratis-GPU).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-04-26 13:24:47 +02:00

203 KiB