Grant: Discovery Development Grant

Digital music catalogs have been growing in volume, due to advances in technologies, including data compression, information retrieval, etc. Music Information Retrieval (MIR) is an emerging area that is engaged in seeking algorithmic approaches to manage large music catalogs. Currently, music catalogs are commonly accessed through tags, such as genres, styles, artists, etc. and it is essential to assign correct tags to music pieces. In this proposed program, my research will center on automatic genre classification in MIR. (1) My first task will extend our current works, which have applied association analysis to verify tag annotation and classify music genres. The weakness is that we obtained our results from some fixed time durations of music pieces, e.g., the middle 30 seconds. I believe that including more time durations would improve genre classification. We plan to not only conduct association analysis within single time durations but also apply sequential association analysis to all durations. It is conjectured that, in addition to the relationship among content-based features within individual time durations, there is also some relationship among them from different durations. Such relationships would help us improve genre classification. (2) If we view the data points in a music piece, our intuition would tell us that those data points are not totally randomly spaced and yet they are not totally regular. The former would mean noise while the latter would correspond to some plain sound. I therefore propose to measure the entropy of data points in a music piece and believe that a music genre should have its own unique entropy, which would help us distinguish music genres. We plan to introduce entropy into analyzing music genre classification, where the music pieces in a genre would be represented by a vector of entropy-based information gains and different genres would have different vectors, distinguishing one from another. (3) Segmentation in music in a long-standing problem in MIR. It is believed that introducing segmentation into analyzing the data points in a music piece, i.e., segmenting it into a set of continuous time durations such that dramatic events are captured between the durations, could lead to better understanding of music data. We have done works on perceptual segmentation and introduced metrics to indicate why a segmentation is better. We will continue our work by proposing some other evaluation metrics. As an example, with (2) in hand, our search for segmentation points in a music piece could be guided by the entropy in music. Once a segmentation is obtained, it could be examined and verified by (1), (2) or some other classification methods. My objective in the new program is to further investigate the segmentation problem in MIR through novel computational means. The work in this proposed program can be easily adapted and extended to style, mood, and artist classification in music.