Classification of non-speech signals with convolutional neural network based approaches

The interest on auto-classification and recognition of different types of data, has increased with development of machine-learning algorithms every passing year. The classification of non-speech sounds and biomedical signals taking part at these interest fields are difficult because of background sound and noise. In the literature, such classification problems are generally tried to be overcome by methods using acoustic features and traditional classification algorithms In this study, the classification problem of non-speech sounds and biomedical signals are tackled with Convolutional Neural Networks (CNN) based approaches, which have become popular recently and are a subfield of deep learning. In this context, the experimental studies are realized on public datasets including lung sounds, heart sounds, EMG signals and environmental sounds. In the proposed methods, studies are generally conducted with deep features based and transfer learning based approaches. It is shown that deep feature-based approaches give better the classification performance than the transfer learning-based approaches and the other conventional methods using same dataset. Besides, at the last part of the thesis, an CNN model whose architecture can be arranged by user and trained with environmental sounds, is proposed to extract deep features instead of using CNN models, which is pre-trained with big image data. It is concluded that this model presents better classification performance than deep feature extraction approach with pre-trained CNN models and other methods using same dataset for the classification of environmental sounds.