SEPARACIÓN DE INSTRUMENTOS MUSICALES CON REDES NEURONALES PROFUNDAS PARA MÚSICA CLÁSICA
Fecha
2023-12-19
Autores
Título de la revista
ISSN de la revista
Título del volumen
Editor
Resumen
[ES] El presente Trabajo de Fin de Grado (TFG) se centra en la SSS en música clásica mediante el uso de redes
neuronales profundas. Para ello se ha llevado a cabo el estudio y análisis de dos algoritmos de redes
neuronales profundas, Open-Unmix y Demucs. Con este fin, se ha preparado un conjunto de datos para
entrenar dos modelos, con el propósito de lograr la separación de fuentes musicales en archivos de
cuartetos de cuerda: violín, viola, violonchelo y bajo. Para permitir a los usuarios realizar separaciones
utilizando estos modelos, se ha desarrollado una aplicación en Python que toma como entrada un archivo
musical y proporciona como salida las fuentes separadas mediante ambos modelos. Finalmente, para una
evaluación objetiva de la calidad de separación obtenida con cada modelo, se ha creado un script de
evaluación que compara las fuentes separadas por los modelos respecto a la separación perfecta, usando las métricas comúnmente empleadas en la bibliografía. Este enfoque permite analizar y comparar el rendimiento de OpenUnmix y Demucs en la tarea de separación de fuentes musicales.
[EN] This Final Degree Project (TFG) focuses on SSS in classical music using deep neural networks. For this purpose, the study and analysis of two deep neural network algorithms, Open-Unmix and Demucs, has been carried out. To this end, a dataset has been prepared to train two models, with the purpose of achieving the separation of musical sources in string quartet files: violin, viola, cello and bass. To allow users to perform separations using these models, a Python application has been developed that takes as input a music file and provides as output the separated sources using both models. Finally, for an objective evaluation of the separation quality obtained with each model, an evaluation script has been reated that compares the sources separated by the models with respect to perfect separation, using the metrics commonly employed in the literature. This approach allows to analyze and compare the performance of OpenUnmix and Demucs in the musical source separation task.
[EN] This Final Degree Project (TFG) focuses on SSS in classical music using deep neural networks. For this purpose, the study and analysis of two deep neural network algorithms, Open-Unmix and Demucs, has been carried out. To this end, a dataset has been prepared to train two models, with the purpose of achieving the separation of musical sources in string quartet files: violin, viola, cello and bass. To allow users to perform separations using these models, a Python application has been developed that takes as input a music file and provides as output the separated sources using both models. Finally, for an objective evaluation of the separation quality obtained with each model, an evaluation script has been reated that compares the sources separated by the models with respect to perfect separation, using the metrics commonly employed in the literature. This approach allows to analyze and compare the performance of OpenUnmix and Demucs in the musical source separation task.
Descripción
Palabras clave
Sistemas de telecomunicación