Use this link to cite:
https://hdl.handle.net/2183/47693 Procesamiento de audio para la separación de fuentes musicales mediante técnicas supervisadas y no supervisadas
Loading...
Identifiers
Publication date
Authors
Meijide García, Clara
Advisors
Other responsabilities
Universidade da Coruña. Facultade de Informática
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen]: Este trabajo aborda la separación de señales individuales en mezclas de audio, como la descomposición de canciones en percusión, voz, bajo y otros elementos, utilizando distintos enfoques metodológicos. Inicialmente, se implementó un filtro adaptativo basado en el algoritmo Least Mean Square (LMS), un método supervisado que ajusta iterativamente los parámetros del filtro para minimizar el error entre la señal estimada y la señal objetivo. Aunque LMS no es una técnica de separación ciega de fuentes, sirvió para explorar la optimización de parámetros en un entorno controlado y supervisado, proporcionando una base técnica sólida. Posteriormente, se desarrolló un modelo no supervisado utilizando el Análisis de Componentes Independientes (ICA). Este enfoque asume que las fuentes en la mezcla son estadísticamente independientes y se mezclan de manera lineal, permitiendo la separación sin necesidad de información previa sobre las señales individuales. El uso de ICA es más adecuado para escenarios reales donde no se dispone de datos de entrenamiento, ya que se basa en la independencia estadística para separar las fuentes de manera efectiva. A lo largo de este trabajo, se experimentó con diferentes parámetros en cada uno de los métodos, evaluando su impacto en la calidad de la separación. Los resultados obtenidos permitieron comparar la efectividad de cada técnica, destacando sus fortalezas y limitaciones en distintos escenarios de aplicación. Este trabajo contribuye al avance del procesamiento de audio, ofreciendo una visión técnica y comparativa de las metodologías modernas para la separación de señales en mezclas de audio.
[Abstract]: This work addresses the separation of individual signals in audio mixtures, such as the decomposition of songs into percussion, vocals, bass, and other elements, using different methodological approaches. Initially, an adaptive filter based on the Least Mean Square (LMS) algorithm was implemented. This supervised method iteratively adjusts the filter parameters to minimize the error between the estimated signal and the target signal. Although LMS is not a blind source separation technique, it served to explore parameter optimization in a controlled and supervised environment, providing a solid technical foundation. Subsequently, an unsupervised model was developed using Independent Component Analysis (ICA). This approach assumes that the sources in the mixture are statistically independent and are linearly mixed, allowing separation without prior knowledge of the individual signals. The use of ICA is more suitable for real-world scenarios where training data is unavailable, as it relies on statistical independence to effectively separate the sources. Throughout this work, experimentation was conducted with different parameters in each of the methods, evaluating their impact on the quality of separation. The results obtained allowed for a comparison of the effectiveness of each technique, highlighting their strengths and limitations in different application scenarios. This work contributes to the advancement of audio processing, offering a technical and comparative perspective on modern methodologies for signal separation in audio mixtures.
[Abstract]: This work addresses the separation of individual signals in audio mixtures, such as the decomposition of songs into percussion, vocals, bass, and other elements, using different methodological approaches. Initially, an adaptive filter based on the Least Mean Square (LMS) algorithm was implemented. This supervised method iteratively adjusts the filter parameters to minimize the error between the estimated signal and the target signal. Although LMS is not a blind source separation technique, it served to explore parameter optimization in a controlled and supervised environment, providing a solid technical foundation. Subsequently, an unsupervised model was developed using Independent Component Analysis (ICA). This approach assumes that the sources in the mixture are statistically independent and are linearly mixed, allowing separation without prior knowledge of the individual signals. The use of ICA is more suitable for real-world scenarios where training data is unavailable, as it relies on statistical independence to effectively separate the sources. Throughout this work, experimentation was conducted with different parameters in each of the methods, evaluating their impact on the quality of separation. The results obtained allowed for a comparison of the effectiveness of each technique, highlighting their strengths and limitations in different application scenarios. This work contributes to the advancement of audio processing, offering a technical and comparative perspective on modern methodologies for signal separation in audio mixtures.
Description
Keywords
Separación de señales Media de Mínimos Cuadrados (LMS) Análisis de Componentes Independientes (ICA) Procesamiento de audio Separación ciega de fuentes Mezcla de señales Filtrado adaptativo Aprendizaje supervisado y no supervisado Signal separation Least Mean Square (LMS) Independent Component Analysis (ICA) Audio processing Blind source separation Signal mixing Adaptive filtering Supervised and Unsupervised learning
Editor version
Rights
Attribution 4.0 International








