Similarity Based Music Retrieval Systems: A Hybrid Approach

Wide-scale distribution of music over the Internet has changed the requirements of the consumer who is exposed to more music than ever before. The consequences of the new era for musical distribution are twofold: Consumers require increasingly personal recommendations for purchases as the potential catalogue expands and Personal musical Libraries of an unprecedented size are becoming the norm as the content becomes increasingly available; current indexing and retrieval systems are inadequate for these tasks. It is in the interests of both the music retail industry and consumers for new methods of indexing and retrieval of music to be developed. Music retrieval is a non-trivial problem requiring some measure of similarity between audio data to be established which is analogous to the hidden similarities in music perceived by humans. This project presents a critical account of the current research into music retrieval systems characterised as approaches from the contrasting Psychological and Physical domains. Psychological approaches expand on the current infrastructure for music classification by human-generated metadata, whilst Physical approaches seek to define music in terms of the features of the audio-data. An intelligent-system is proposed which learns to synthesise playHsts from humangenerated training-data through a novel combination of Sardnet clustering, Genetic- Algorithm searching and Radial Basis Function feedback