ESPHome lets makers place active wake-word models in faster internal memory, reducing the processing time needed to listen for a wake word. The option starts off and falls back to the existing memory arrangement when there is not enough internal space.
In contributor embe’s test on an ESP32-WROVER with one 16 kHz model, each processing call fell from about 13.3 milliseconds to 3.7 milliseconds. A separate playback test reduced wake-word processing from about 64% to 33% of one core. Those measurements describe that device and model.
Only active models use the extra memory, including models loaded while the device is running. Disabling a model releases it again. The change speeds model processing; it does not establish better recognition accuracy. This is embe’s first contribution to ESPHome, reviewed by J. Nick Koston.