| With the rapid development of modern science technology, music information retrieval, a newly developing field of studying, has become a hot research subject across the world. As a pivotal supplementary mean of music information retrieval, the study of automatic music classification shows great significance both in theory and in practice.China owns a vast area of territory and thriving historical cultures, which give birth to the unique Chinese folk music. It has strong regional and ethical characteristics that differ from other countries'music in various aspects. The study on contents and instrumental style of Chinese folk music can be extremely beneficial either in exploring variation of musical culture or historical music restoration. In this dissertation, we present a series of studies on the automatic classification of Chinese folk music by means of three approaches—feature analysis and extraction, design of the classifier and feature selection.In order to build up the Chinese folk music database, we collected 642 pieces of traditional music which belongs to 8 different categories as Guqin, Guzheng, Yangqin, Erhu, Dizi, Pipa, Nanyin and Suona. Then we sliced them on 10s and 20s scale respectively so two musical datasets were obtained, namely Folk10s and Folk20s, which are the major source of experimental data.Generally, automatic classification of music can be regarded as a two-stage procedure—feature extraction and categorization. Considering 4 aspects of tone, loudness, timbre and rhythm, we perform an in-depth study into characteristics of each type of folk music before picking out 14 representative features (70 dimensions in sum) which were used in classification experiments. Then the evaluation of five common classifiers (BP Neural Network, Naive Bayes, K-nearest neighbors, Decision Tree and Support Vector Machine) was performed on FolklOs and Folk20s datasets respectively. According to the result, SVM turned out to be the most effective method when coping with the classification of Chinese folk music.Besides, we also compared the performance of different acoustic feature sets and combination of them using SVM on our data sets, together with the evaluation of impacts on classification accuracy when using several types of core functions.Though SVM represented competitive performance in classification experiments, inevitably, there were still some useless and redundant features in the feature sets. In order to combat this situation and enhance the performance of the classifier, we firstly perform feature selection on Folk10s dataset using 3 three filter methods, namely ReliefF, ReliefF-PCA and ReliefF-CA, then put forward two new filter-wrapper methods—Heuristic Forward Search (HFS) and Heuristic Backward Search (HBS), which showed excellent performance in the following feature selection experiments. The further examinations on two non-musical datasets using HFS and HBS demonstrated their effectiveness and robustness.As a special form of traditional music, Chinese folk music brings us a strong sense of beauty and marvelous artistic value. The study of content-based music classification can greatly benefits the further folk music retrieval and other processing techniques such as music restoration and digital reproduction,so engaging in it turns out to be essential and significant. |
