A comparative study of pitch estimation algorithms for monophonic and polyphonic music signals

Abstract

Pitch extraction has been a prevalent subject in numerous areas of research ever since the era of computers. Pitch detection is a fundamental problem in a number of fields, such as Music Information Retrieval (MIR) and automated score writing. It has become beneficial in new zones lately, such as computer games. It is also used in wave to MIDI converters, such as "Digital Ear" [Epinoisis Software].

The extraction of music features has a history of some 50 years. Yet recent technologies couldn't make it to a preferred level of precision and robustness. Most of the technologies work well for clean, noiseless single tones. However the algorithms fail miserably when it comes to multi pitch signals.

This dissertation presents some of the well-known pitch estimation algorithms in monophonic as well as polyphonic music. The algorithms were implemented using MATLAB and was tested initially with the synthesized music signals and later with original instrumental music pieces. A comparison of the algorithms is also discussed based on the test results. An outlook into some of the pitch based instrument classifiers is also presented.