Feature extraction with convolutional restricted boltzmann machine for audio classification

Feature extraction is a crucial part for a large number of audio tasks. Researchers have extracted audio features in multiple ways, among which some most recent methods are based on the hidden layer of a trained neutral network. In this paper, we present a system which can automatically extract features from unlabeled audio data, and then the features of extracted from the system are used for audio classification task. Ourfeature extraction scheme makes use of a convolutional restricted Boltzmann machine (CRBM), instead of those using restricted Boltzmann machines (RB-M). By using features extracted from CRBM, we can achieve about 7% accuracy improvement consistently over than the RBM-based features on the TI-Digits dataset for audio classification. We also combine the well-known MFCC features and the CRBM-based features in the form of a linear combination. In our experiments, this feature combining the two methods performs better than both features alone.