Mel Power Spectrogram Approximation By Tiny Neural Networks for Home Appliances Classification

Mel power spectrogram has been widely deployed in audio preprocessing for signal conditioning and feature extraction. Between many available, one of the most used software libraries is named Librosa. This paper investigates how the Mel power spectrogram algorithm can be approximated by tiny neural networks. The proposed approach would allow to build an end-to-end multitask deep neural pipeline coded by using Keras layers. By connecting proposed approximator to a classifier, it is possible to estimate the Mel spectrogram and achieve adequate classification accuracies. These approaches have been tested with WHITED, COOLL and PLAID datasets for electric current classification of home appliances. The approximator achieved on the WHITED dataset the mean square error of 1.46 e-03, on COOLL was 0.396 and 1.84 on PLAID. The classification of the home appliance types achieved best accuracy of 96.43% for 7 classes on WHITED, 95.63% for 5 classes on COOLL and 76.61% for 6 classes on PLAID. The complexity of the tiny networks has been set and profiled with respect to the limited memory requirements of the off-the-shelf ARM Cortex M4 and M7 microcontrollers. The deployment on hardware boards has been conducted too. The inference time on the ARM Cortex M4 was 5092ms, while on M7 was 814.7 ms.