Evolución de sistema multiclasificador basado en programación genética cartesiana para estimación multitono de audio de piano

Abstract

Multi-Pitch Estimation, or multiple fundamental frequency estimation, is the process of extracting the musical notation (pitches) from a given acoustic signal. Multi-Pitch Estimation is one of the tasks that belong to Content-based Music Information Retrieval. Among all musical instruments, piano is one of the most popular worldwide, and one of the most complex concerning pitch variety and number of simultaneous notes. These are the main reasons that motivate us to research on Multi-Pitch Estimation of piano sounds. The problem of Multi-Pitch Estimation is addressed as classification problem, with the main objective being to find the musical notes that are present in an observed sound signal. This problem is tackled using one of the most prominent and recent methodologies of the family of Evolutionary Algorithms -- Cartesian Genetic Programming. This thesis presents a novel approach to the problem of Pitch Estimation, using Cartesian Genetic Programming. This thesis describes a research that started with the need of developing a generic toolbox for Matlab, capable of aiding users to encode problems using Cartesian Genetic Programming. Using a small step iterative approach, we tackled the problem of Multi-Pitch Estimation, starting with a first approach to the problem for piano sounds. The system has undergone several improvements in order to become more accurate, flexible and faster. Taking advantage of the multi-classifier architecture we reached the real time performance using a several core processor with the classifiers distributed. We also extended the technique to other musical instruments like guitar and we showed its feasibility with the supported of experiments and results.