Abstract
A music piece usually consists of a sequence of sound events which represent both short-term and long-term temporal information. However, in the task of automatic music genre classification, most text-categorization-based approaches only capture temporal local dependencies (e.g., unigram and bigram-based occurrence statistics) to represent music contents. In this dissertation, we propose to use time constrained sequential patterns (TSPs) as effective features for music genre classification. First of all, an automatic language identification technique is performed to tokenize each music piece into a sequence of hidden Markov model indices. Then TSP mining is applied to music sequences to discover genre-specific TSPs, followed by the computation of occurrence frequencies of TSPs in each music piece. Finally, these occurrence frequencies are feed into support vector machines (SVMs) to perform the classification task. Experiments conducted on two widely used datasets, GTZAN and ISMIR2004Genre, show that the proposed method can discover more discriminative temporal structures and achieve a better recognition accuracy than the unigram and bigram-based statistical approach. In addition, we also propose another music genre/mood classification system which combines both short-term frame based timbre features and the long-term modulation spectral analysis of timbre features for SVMs. This proposed system won the first place of the MIREX 2011 music mood classification task. In our submission, we performed the modulation spectral analysis on short-term timbre features to extract long-term modulation features. However, two operations in this analysis are likely to smooth out useful modulation information, which may degrade the classification performance. The first one is to take the averaging of modulation spectrograms extracted from texture windows (each of which is composed of timbre features extracted from hundreds of frames) to create a representative modulation spectrogram for a music clip. The second one is to compute the mean and standard deviation of modulation spectral contrast/valley matrices (these two matrices are computed from the representative modulation spectrogram) to obtain a compact feature vector for a music clip. To avoid smoothing out modulation information, in this dissertation, we propose the use of a two-dimensional representation of acoustic frequency and modulation frequncy to compute joint frequency features. These joint frequency features, including acoustic-modulation spectral contrast/valley (AMSC/AMSV), flatness measure and crest measure (AMSFM/AMSCM), are then computed from modulation spectra of each joint frequency subband. By combining the proposed features, together with the modulation spectral analysis of MFCC, and statistical descriptors of short-term timbre features, this new set of features outperforms our MIREX 2011 submission on four other genre/mood datasets.