Deep learning based urban sound classification and ambulance siren detector using spectrogram

Abstract

An efficient sound classification algorithm can benefit a multitude of applications involving home, wildlife, and residential surveillance, traffic regulation, medical monitoring etc. An important application is to identify and notify the ambulance siren sound amidst a noisy environment. This work develops an efficient and less complex architecture for siren detection and urban sound classification using different sound to image transformation methods viz: Mel-spectrogram, Scalogram and Fourier decomposition method(FDM). Effects of augmentation and pre-processing techniques on the efficacy of the developed architecture against pre-trained models is analyzed. The performance of the proposed algorithm is tested on Urbansound8K dataset for urban sound classification, and a multi -source dataset for ambulance siren detector. The CNN proposed here gives an accuracy of 89.66% for the former case and 99.35% for the latter.