Coding Mel Spectrogram using Keras and Tensorflow for Home Appliances Tiny Classification

Mel power spectrogram has been extensively used as audio pre-processing for both feature extraction and transformation. Between many, one of the most used libraries is Librosa. In this paper, we prove that the Mel power spectrogram processing algorithm can be coded using Keras and Tensorflow software primitives featuring several statically initialized or trainable hyper-parameters. That approach would allow to achieve a self-contained end to end deep learning framework dependent implementation not possible with other software libraries. Moreover, it can be either initiable with pre-computed parameters or let them be trainable as part of the same multitask model as any subsequent machine learning approach. The pipeline has been tested on COOLL, WHITED and PLAID datasets for electric current classification of home appliances. The difference in accuracy was +3.28% on COOLL, + 1.56% on WHITED and −0.16% on PLAID with respect to reference work from same authors. The complexity of the proposed approach has been estimated in term of multiplied and accumulated operations. It resulted be 48,821 MACC/s for an estimated number of cycles/s of 439,389 for STM32L4 and 292,926 for STM32H7.