Melody-Based Approaches in Music Retrieval and Recommendation Systems

Abstract

Analysis and description of digital audio, being one of the fundamental information sources in the modern world, becomes surprisingly complex when it comes to music information retrieval. In addition to the numerous known classic problems, such as pitch and rhythm extraction, tempo analysis, audio key detection or chord estimation, there exist further considerable difficulties in higher-level analysis involving e.g. emotional content or genre categorization. These issues are of utmost interest for the rapidly growing music industry and new e-commerce solutions which necessitate the need for precise estimation of users' musical preferences. Out of the many proposed approaches, those based on modelling the social context and behavior of a user via collaborative filtering techniques, although potentially effective, do not offer real insight into the decision-making process. The content-based methods, on the other hand, enable to build refined models of users' preferences with adjustable weights attributed to different musical elements and varied description levels. This chapter presents an evaluation of the latest efforts and development in fields related to music information retrieval, as well as a study of melody based approaches for use with recommendation systems. The evaluation concentrates on algorithms extracting audio features relevant to the idea of similarity analysis and recommendation proposals. Only content-based audio analysis approaches are taken into account. Tasks such as audio key detection, chord estimation, tempo analysis or genre classification are described, latest achievements and algorithms are presented and ways in which they can be effectively used in recommendation systems are estimated. The concept of audio melody extraction is explored and the authors' own research results in this subject are presented. Possible usage scenarios of melody analysis for music recommendation services are evaluated. Application in conjunction with query by humming and query by tapping systems is analysed as an example of a complete recommender solution. © Springer International Publishing Switzerland 2013.