A typology for voice and music signals

Abstract

With the high increase in the availability of digital music, it has become of interest to automatically query a database of musical pieces. At the same time, a feasible solution of this objective gives us an insight into how humans perceive and classify music. In this research, we discuss our approach to classify music into four categories: pop, classical, country and jazz. Songs are collected in wave format. We randomly chose five 10-second clips from different parts of a song. We discussed two families of features: wavelet features and time-based features. These features are capable of capturing the information of energy and time of voice signal. Instead of using traditional Mel-Frequency Cepstral Coefficients (MFCC)[7] methods, which are widely used in audio classification and music classification, we incorporate the features in statistical classification methods such as LDA, QDA and tree. Finally, we attempted to create an adaptive tree approach for classification. In this research, 130 songs are collected. Pop songs are collected in 4 languages, English, Chinese, Spanish and French. A cross validation method is used to compute the proportion of correctly classified songs. It is shown that the tree method has a proportion of correct classification equal 0.80 when pop and country are considered as one category.