Effective music retrieval by sequential Pattern-Based alignment

Due to the rapid growth of music data, how to effectively and efficiently retrieve the interested music piece has been an attractive issue in recent years. In traditional music retrieval systems, the most popular way is to retrieve the music piece by matching query terms and music profiles like file name, artist and so on. However, this type of music retrieval systems suffers from problem of semantic gap. To aim at this problem, in this paper, we propose a new method named Pattern-Based Music Retrieval named PBMR that exploits temporal continuities of acoustical content to represent the musical features. That is, a music piece in this work is first converted into a pattern string by two-stage clustering. Thereupon the similarity between two music pattern strings is calculated by alignment-like algorithm. The experimental evaluations show that our proposed perceptual patterns are sensitive for listening and temporal continuities are helpful to identifying the similarities between music pieces.