Speech-music discrimination: A deep learning perspective

This paper is a study of the problem of speech-music discrimination from a deep learning perspective. We experiment with two feature extraction schemes and investigate how network depth and RBM size affect the classification performance on publicly available datasets and on large amounts of audio data from video-sharing sites, without placing restrictions on the recording conditions. The main building block of our deep networks is the Restricted Boltzmann Machine (RBM) with binary, stochastic units. The stack of RBMs is pre-trained in a layer-wise mode and, subsequently, a fine-tuning stage trains the deep network as a whole with back-propagation. The proposed approach indicates that deep architectures can serve as strong classifiers for the broad binary problem of speech vs music, with satisfactory generalization performance.