The Study Of The Automatic Classification Of Chinese Folk Songs By Regional Style

Music classification is an important topic in the area of music information retrieval. Audio is an important part of the multimedia. As repid increase in audio data, how to automatically manage these data has become a prominent issue. Especially to the wide variety of music data, people demand a rapid and efficient method for the management of their classification (according to different styles or singers, etc.). This requires an efficient automatic classification technology to collate audio data, and it can serve audio search or related analysis.China is a far-flung country with 56 different nationalities. In China, different nations inhabit historically in accordance with geographical residence. Each nation not only has its own customs, and also has its own distinctive language and the special style of music. The Han nationality is the most populous nation, dwelling in vast areas, but different dialects and different styles of folk music in accordance with different region. Research on regional character and regional style of Chinese folk song is a very important and meaningful work in the fields of musicology and music information retrieval. In this dissertation, we present a study on automatic classification of Chinese folk songs by regional style using signal processing and machine learning technologies. Our research mainly focus on three aspects: feature extraction, classifier selection and feature selection.In this dissertation, we built two Chinese folk song databases, one is original Chinese folk song database, the other is created Chinese folk song database. There are folk songs from ten regions in each database. According to segmented length of each clip, we got four datasets: Original10s, Creative10s, Creative30s and Creative60s. All experiments in this dissertation are based on these four datasets.There are usually two phases in Automatic music classification: feature extraction and classification. We extract 16 features among which 11 are timbre related, 3 are rhythm related, 2 are loudness related from each folk song clip, totally 74 dimensions. Timbre related features include Mel-frequency cepstral coefficients (MFCC), Liner predictive coefficients (LPC),spectral centroid, spectral flux, zero crossing rate and so on. Rhythm related features include strongest frequency, strength of strongest frequency and beat sum. Loudness related features include root mean square and fraction of low energy windows. We compared the performance of five classifiers which are often used in audio classification include Na?ve bayes classifier, Fisher Linear Discriminant Analysis, K nearest neighbors, Neural Network and support vector machine on extracted feature vectors. The results show that support vector machine classification performance has a certain advantage. In order to further improve the classification accuracy, we use majority voting algorithm to combine multiple classifiers, and propose a new classifier ensemble method based on optimization base classifiers.In order to evaluate the performance of each feature set in regional style based Chinese folk song classification, we did experiments on different feature sets of Original10s dataset and Creative30s dataset using support vector machine classifier. Feature selection is a key technology in pattern recognition which can reduce the classifier model training time, improve the classification accuracy, and discover the most important feature. In this dissertation, we use filter feature selection algorithms such as ReliefF and Fisher rule, and wrapper feature selection algorithm such as sequential forward selection with nearest neighbor classifier to select important features and reduce feature vector dimensions on original folk song dataset Original10s. At last, we propose an efficient Active Selection feature selection algorithm based on a mixture feature selection model.Audio classification technology is an important supplementary means for the audio retrieval and other audio processing. Through content based music classification, it brings the convenience to music retrieval or related analysis. Chinese folk song is a special music form which contains abundant regional styles and high artistic value. Therefore, the study of the classification of Chinese folk songs by regional style is a very important and meaningful work.