Improvement of Automatic Genre Classification System for Traditional Malaysian Music Using Beat Features

• Main Author: Mohd. Norowi, Noris
• Format: Thesis
• Language: English
• Published: 2007
• Online Access: http://psasir.upm.edu.my/id/eprint/5205/1/FSKTM_2007_7a.pdf; http://psasir.upm.edu.my/id/eprint/5205/
• Institution: Universiti Putra Malaysia

The increase in processing power and storage of computer has resulted in the growth of digital musical files, which demands some form of organization such as classification of the files. Typically, manual classification is used but it is expensive both in terms of time and money. One alternative solution is to automate musical genre classification. Existing systems have been developed to classify Western musical genres such as pop, rock and classical. However, adapting these systems for traditional Malay music is difficult due to the differences in musical structures and modes. In general, the musical structure of many genres in traditional Malay music is rhythmic and repetitive, which is different than Western music. This study investigates the effects of factors and audio feature set combinations towards the classification of traditional Malay musical genres. Ten traditional Malay musical genres are introduced in this study: Dikir Barat, Etnik Sabah, Gamelan, Ghazal, Inang, Joget, Keroncong, Tumbuk Kalang, Wayang Kulit and Zapin. The study is conducted in three phases. The first phase investigates the factors affecting classification of traditional Malay music: dataset size, track length, track location, number of cross-validation folds, and classifier. The second phase investigates the effect of feature set combinations on the classification result of traditional Malay music. The combinations are STFT, MFCC, STFT and Beat, MFCC and Beat, and STFT, MFCC and Beat. Following this, an automated classification system is developed and named MAGCLAST (Musical Analysis and Genre CLAssification System for Traditional Malay Music). The performance of MAGCLAST against three groups human (expert, trained and untrained) is tested in the final phase of the study. Results show that its classification at 66.3% is comparable to MARSYAS (61%) and trained human (70.6%). Interestingly, MAGCLAST also outperforms classification by average Malaysians, suggesting that an automated system for classifying traditional Malay music is certainly needed. Additionally, a small-scale study on human classification behaviour is also done to understand the factors that affect classification. It is hoped that the information could be exploited to enhance existing automated genre classification system in the near future.