voice-models

Custom wake word models for ESPHome / microWakeWord.

my_wake_word

Trained with microWakeWord locally on Windows + NVIDIA GPU.

Performance (quantized streaming TFLite):

• Cutoff 0.93: 2% miss rate, ~0.4 false positives/hour
• Cutoff 0.99: 8% miss rate, ~0.2 false positives/hour

ESPHome config

micro_wake_word:
  models:
    - model: https://raw.githubusercontent.com/BareMetalTinker/voice-models/main/my_wake_word/manifest.json