A new mechanism of similarity evaluation for content-based music information retrieval

Abstract

Content-based music information retrieval (CBMIR) has rapidly become a research focus for the areas of computer science, information retrieval, signal processing, audio processing and pattern recognition. Feature selection, representation and matching mechanism play the crucial roles in CBMIR. A new mechanism of similarity comparison has been proposed in this paper. Melody feature is represented in pitch interval based on physics and perception characteristic of music, which averts the effect by gross differences in key or tempo. The Longest matched subsequences (LMS) algorithm is proposed to obtain the most matched portions from two music pieces, according to local similarity between elements. A compound criterion of similarity evaluation is established, with both matched proportion and distance of matched subsequences being considered. The feasibility and validity of the model is verified by the experiment with a music feature database containing hundreds of songs.