A comparison of features for speech, music discrimination

Several approaches have previously been taken to the problem of discriminating between speech and music signals. These have used different features as the input to the classifier and have tested and trained on different material. In this paper we examine the discrimination achieved by several different features using common training and test sets and the same classifier. The database assembled for these tests includes speech from thirteen languages and music from all over the world. In each case the distributions in the feature space were modelled by a Gaussian mixture model. Experiments were carried out on four types of feature, amplitude, cepstra, pitch and zero-crossings. In each case the derivative of the feature was also used and found to improve performance. The best performance resulted from using the cepstra and delta cepstra which gave an equal error rate (EER) of 1.28. This was closely followed by normalised amplitude and delta amplitude. This however used a much less complex model. The pitch and delta pitch gave an EER of 4% which was better than the zero-crossing which produced an EER of 6%.