Features for Audio and Music Classification.pdfVIP

Features for Audio and Music Classification Martin F. McKinney Philips Research Laboratories Prof. Holstlaan 4 (WY82) 5656 AA Eindhoven, The Netherlands martin.mckinney@ Jeroen Breebaart Philips Research Laboratories Prof. Holstlaan 4 (WY82) 5656 AA Eindhoven, The Netherlands jeroen.breebaart@ Abstract Four audio feature sets are evaluated in their ability to classify five general audio classes and seven pop- ular music genres. The feature sets include low-level signal properties, mel-frequency spectral coefficients, and two new sets based on perceptual models of hear- ing. The temporal behavior of the features is ana- lyzed and parameterized and these parameters are in- cluded as additional features. Using a standard Gaus- sian framework for classification, results show that the temporal behavior of features is important for both music and audio classification. In addition, classifica- tion is better, on average, if based on features from models of auditory perception rather than on standard features. 1 Introduction Developments in Internet and broadcast technology enable users to enjoy large amounts of multimedia content. With this rapidly increasing amount of data, users require automatic methods to filter, process and store incoming data. Some of these functions will be aided by attached metadata, which pro- vide information about the content. However, due to the fact that metadata are not always provided, and because local pro- cessing power has increased tremendously, interest in local au- tomatic multimedia analysis has increased. A major challenge in this field is the automatic classification of audio and mu- sic (Wold et al., 1996; Spina Zue, 1997; Scheirer Slaney, 1997; Scheirer, 1998; Wang et al., 2000; Zhang Kuo, 2001; Li et al., 2001; Tzanetakis Cook, 2002). Most audio classification systems combine two processing stages: feature extraction followed by classification. A vari- ety of signal features have been used for this purpose, includ- ing low-le