Environmental sound classification using One/Few Shot Learning with Siamese Networks

The inability to tolerate everyday sounds, also known as Decreased Sound Tolerance (DST), has proven to be one the most prevalent issues in Autism Spectrum Disorder (ASD). While advanced Neural Networks have shown promising results in classifying environmental sounds, those conventional classification models rely on sound classes that were used in the training process. In DST, the list of aversive sound classes may be unique and different for each person, and training a conventional classification model that can classify all possible aversive sound classes is not feasible. Hence, a classification approach that works beyond this limitation is required. In this thesis, the idea of One/Few Shot Learning for environmental sound classification is explored. This model can classify a given sound by having one or very few samples of that class. As a part of this research, different aspects of the model are optimized and a state-of-the-art model is developed.