Optimizing Melodic Extraction Algorithm for Jazz Guitar Recordings Using Genetic Algorithms

Extraction of the main melody of a musical piece is a preliminary step in the process of transcribing the piece. Automatic melodic extraction is the task of computationally extracting what a human listener would perceive as the main melody of a polyphonic recording. Several melodic extraction systems have been proposed. However, such systems normally require a number of parameters to be manually tuned in order to accurately perform melody extraction in different contexts, i.e. instruments combinations. In this study, we propose a methodology for automatically optimizing some relevant parameters of a melody extraction algorithm using genetic algorithms. We simultaneously obtained both MIDI and audio recordings of jazz standards, and we collected commercial audio recordings extracted from jazz guitar CDs. Based on the MIDI recordings as ground truth, two different instrument settings are compared (Jazz trio and quartet), as well as different audio mixing of the melody with respect to the accompaniment track. We show that, compared to using the default parameters, the overall accuracy of the melody extraction with the optimized parameters is improved.