This repository is a companion to the Sound Recognition with Edge Impulse tutorial on the Microchip Developer website; it contains the firmware to classify between vacuum cleaner sound and background noise on a SAME54 Curiosity Ultra development board + WM8904 audio daughterboard, as well as some code for creating a custom processing block with Edge Impulse.
Measured with -O2 compiler optimizations and 120MHz Clock
- 144kB Flash
- 76kB RAM
- 98ms Inference time
- 96.9% Test set accuracy (see section on Audio Dataset)
A dataset for vacuum cleaner detection can be downloaded from the releases page. The data is a selected subset of the DEMAND and MS-SNSD datasets, with the vacuum cleaner sounds pulled from MS-SNSD, and the background noise segments (babble noise, air conditioner, and a mix of indoor domestic noise) pulled from both DEMAND and MS-SNSD.
Code for adding LogMFE audio features to Edge Impulse Studio is located under the custom-processing-blocks directory.