Novel Approach for Music Search Using Music Contents and Human Perception

Music similarity can be perceived by the listeners in different ways depending on individual preferences to musical parameters. Metadata based similarity measures are used by many websites for music recommendation and retrieval. Metadata attached to music files as text generally include title, albums, genre, artist etc. The main limitation of this method is that it can only work if metadata is available with the music. It may produce wrong results if incorrect metadata is associated with the music file. Present Meta data based systems also lack in analyzing contents. We have conducted experiments with Indian film songs in pairs and triplets to understand listener´s perception of similarity. We analyzed content based parameters such as voice, mood, instruments used, lyrics, rhythms, tune etc and their relative importance in similarity perception to majority of the listeners. We propose to use both metadata and content based parameters. We worked on few parameters and worked with some algorithms in our prototype. Our findings can be useful for music Information Retrieval System and the concept is applicable to all musical forms. Perception based framework for music similarity is an attempt to model present and future needs of music listeners with option of adjustment for various musical parameters from listener perceptive.