microWakeWord library - community-trained wake words
A growing library of community-trained wake word models ready to drop into ESPHome. Every entry is a quantized INT8 TFLite model paired with its JSON manifest, with a synthetic preview so you can hear the word before you download it.
What's in each download
Each download is a ZIP containing the .tflite model, a manifest JSON with the correct probability_cutoff, sliding_window_average_size, feature_step_size, and tensor_arena_size for ESPHome's micro_wake_word component, and a README with a copy-paste YAML snippet.
Quality signals
Models are benchmarked against noise-augmented positive clips plus real background noise and real conversation audio, so the reported recall and false-accepts-per-hour metrics correlate with real-room performance rather than studio-quality test samples.
Trained with Optuna
Optuna-optimized models are flagged with a badge. For each Optuna run, only the top 5 models (lowest false-accepts-per-hour at the highest available recall tier) are published to the community library.
More
This page is an app that requires JavaScript. Enable JavaScript to use the training center and interactive library.