Clasificación de sonidos ambientales usando la transformada wavelet continua y redes neuronales convolucionales

Abstract

This paper proposes a scheme in which a time-frequency representation is first obtained using the continuous wavelet transform (CWT), which has a logarithmic resolution in the frequency domain, like that of the human ear. The development of these environmental sound classification systems is a topic of extensive research due to its application in several fields of science and engineering. Like other classification schemes, they are based on the extraction of specific parameters that are inserted in the classification stage. The CWT is then inserted into a deep learning neural network to carry out the classification task. The evaluation results obtained using several databases such as ESC-50, TUT Acoustic Scene, and SONAM-50 show that the proposed scheme provides a classification performance that is better than that provided by other previously proposed schemes.