Impact of Noise on Different Neural Network Architectures for Environmental Sound Classification

This thesis aims to determine how effective it is to use advanced techniques in Deep Learning for the classification of environmental sounds, focusing mainly on the recognition of human activities, in the presence of background noises of different nature, and highlighting their performance and robustness in different conditions. The study uses four Deep Learning models and an ESC-50 reference dataset, which has been reduced to 19 classes by selecting only domestic and human sounds. Some of the selected models had already been pre-trained with AudioSet, a larger dataset of environmental sounds. The models were exposed to various signal-to-noise ratio (SNR) levels. The noises used during the test phase are as follows: white, red, blue, violet, television, the opening and closing of the door, the bark of a dog, the engine of a car and the rain falling on the window. The results obtained contribute to the understanding of how to deal with the problem of classification of environmental sounds in noisy environments, offering ideas for the development of more reliable and adaptable algorithms.