Análisis comparativo de los métodos REPET+ y UNet para la separación de la voz cantada en una pista musical

• Autor: Ramon Zuta, Jorge Luis
• Formato: tesis de grado
• Fecha de Publicación: 2023
• Institución: Universidad de Lima
• Repositorio: ULIMA-Institucional
• Lenguaje: español
• OAI Identifier: oai:repositorio.ulima.edu.pe:20.500.12724/17755
• Enlace del recurso: https://hdl.handle.net/20.500.12724/17755
• Nivel de acceso: acceso abierto
• Materia: Análisis musical; Procesamiento de datos; Análisis de Fourier; Análisis matemático; Sistemas de procesamiento del habla; Musical analysis; Electronic data processing; Fourier analysis; Mathematical analysis; Speech processing systems; https://purl.org/pe-repo/ocde/ford#2.02.04

Music source separation is the task of isolating the musical phrases played by different instruments recorded individually and arranged together to form a song. Nowadays, several methods have been developed to cover the separation of music sources, which can be classified into supervised and unsupervised learning, however, no research has been developed in which the effectiveness of using different methods together are analyzed , that's the reason the present work seeks to measure the results of the use of two methods, REPET + (unsupervised) and UNet (supervised), jointly and in isolation to separate the music waves produced by a singer and the waves from the instruments. The results show an overall score (SDR) of the methods for vocal separation for the UNet network was 5.38 dB, REPET+ -4.3 dB, -2.55 dB for REPET+ & UNet, -0.38 dB for UNet & REPET+, -6.16 dB for REPET+ & REPET+ and 5.17 dB for UNet & UNet, demonstrating the superiority of the UNet network for the separation of vocal waves compared to the REPET+ method. In addition, the use of the methods together shows a slight improvement in certain evaluation metrics, however, considering all the metrics (SDR, SIR and SAR), it is evident that this leads to a loss of information that results in a low overall score of the solution.