Melody spotting using hidden Markov models

This thesis describes a new method for melody-based retrieval of digitized music stored in a musically-unstructured format. Using a methodology similar to that of a keyword spotting system in speech recognition, we construct a key melody spotting system using hidden Markov models (HMMs). We build small, note-level HMMs and use them to construct a query model capable of rapidly searching through a large music database for a melody, no matter where it is located in a song. We call this approach melody spotting using hidden Markov models (MSHMM).; This thesis represents a novel approach to query by melody, particularly in the realm of continuous format music (e.g., non-MIDI). While this system does not always improve upon the performance of symbolic music information retrieval (MIR) systems, it nevertheless provides a foundation for and proof of this melody spotting concept. We first describe the basic framework of our MSHMM system and the feature sets used in conjunction with it. We then address extending that system to incorporate notions of invariance and error tolerance in both pitch and duration aspects of melody.; This thesis presents necessary music background information and the origination and history of this melody-based retrieval problem. It then describes our basic MSHMM system and its extensions and evaluates the performance of each. Finally, we discuss future research directions stemming from this work and the conclusions to which this system leads us, that hidden Markov model based keyword spotting methods drawn from the speech recognition domain can be successfully adapted to perform melody-based retrieval of digitized music.