Experiments on Speech/Music Discrimination

The problem of speech/music discrimination has become increasingly important as automatic speech recognition system are applied to more real-world multimedia domains. One of the issue in the design of a signal classifier is the selection of an appropriate feature set that captures the temporal and spectral structures of the signal. Many features have been used in speech/music discrimination. The cepstral coefficients and the posterior-based features are two widely used features. Another issue is the selection of a classification algorithm, and a popular approach to speech/music discrimination is to use the decorrelated feature frames to train distribution models to distinguish training data labelled as speech or music [4, 10]. In this report, we test the performance of three kinds of classic features (MFCCs, Dynamic features, and the posterior-based features) and two kinds of classic distribution models (GMM and K-nearest neighbor) in separating speech and music signals. Furthermore, we implement a demo to classify a given audio signal into speech or music category.