A framework for pattern based melody matching for Content Based Music Information Retrieval

Content Based Music Information Retrieval (CBMIR) Systems help the users to find the interesting musical object from a vast collection of musical objects based on the content expressed in terms of musical phrases referred to as repeating patterns in response to query often expressed as a smaller fragment of note sequence. It is crucial to identify repeating patterns, indexing the musical objects based on the patterns and estimate the relevance of music objects to the given query for preparing the ranked list of music objects. This paper discusses development of a framework for pattern based melody matching used to build CBMIR systems. The framework consists of five modules to support the content processing of music objects for multiple tasks. Module-1 deals with extraction of melody track from the music object and representing it as a symbolic note sequence. Alternative representational strategies and their suitability to different scenarios are discussed. Module-2 deals with extraction of approximate repeating patterns from the note sequences representing the music objects to identify semantic features of music object. Module-3 applies document retrieval techniques to transform the music objects into semantic feature space using the approximate patterns identified by the previous module. A pattern base is created to maintain the inverted list of music objects (along with the prominence scores) corresponding to each pattern. Query preprocessing to transform it into a set of query terms followed by query pattern matching with candidate patterns available in the pattern base is implemented in Module-4 of the framework. Finally the Module-5 estimates the matching scores of the music objects/songs if they contain some/all of the query patterns and sort the music objects in the order of their matching scores. Experimentation is conducted on two real world dataset of musical objects: one containing South Indian classical music and the other containing popular movie songs of India. The performance of the framework is estimated in terms of Mean Reciprocal Ranking (MRR) and is found to be satisfactory even for short queries.