Publication: Aplicación de algoritmos combinados de filtrado adaptativo a acústica de salas
Loading...
Identifiers
Publication date
2011
Defense date
2011-06-09
Authors
Advisors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
Las aplicaciones de procesamiento de señales acústicas están cobrando una importancia creciente. La mayoría de aplicaciones de este tipo (como la cancelación de eco acústico, la cancelación de ruido, la dereverberación, la separación y el seguimiento de fuentes acústicas, etc.) requieren la identificación de una (o varias) respuestas al impulso del recinto (RIRs). Estas respuestas pueden variar con el tiempo, por lo que se precisa de esquemas adaptativos para su identificación. La utilización de esquemas adaptativos en escenarios de identificación de respuestas acústicas se ve sujeta a diferentes compromisos, como, p. ej., la conocida relación entre velocidad de convergencia y precisión en estacionario. Varios de estos compromisos se comparten con otras aplicaciones, mientras que otros son específicos del procesamiento de señales acústicas. Entre los diferentes métodos que tratan de aliviar estas limitaciones, destaca la combinación adaptativa de filtros adaptativos debido fundamentalmente a su sencillez, versatilidad y eficacia. En esta Tesis Doctoral se aborda el estudio, diseño, implementación y adecuación de los esquemas de combinación adaptativa para que resulten provechosos y convenientes en aplicaciones de procesamiento de señales acústicas. Para ello, se proponen y analizan esquemas de combinación que ofrecen robustez y un comportamiento adecuado con respecto a las particularidades que presentan las señales acústicas involucradas y las RIRs. De entre los posibles condicionantes y sus potenciales soluciones, en esta Tesis Doctoral se contemplan: - La relación señal a ruido es normalmente desconocida a priori y puede variar. Se han desarrollado dos esquemas de combinación de filtros robustos frente a cambios en dicha relación. - El espectro de las señales acústicas (música y voz) no es plano en frecuencia, lo que ralentiza la convergencia de los filtros adaptativos. Se presenta un algoritmo de combinación en el dominio frecuencial que permite combinar de forma independiente diferentes bandas de frecuencia, obteniendo ganancias debido a que, por lo general, la relación señal a ruido es diferente en cada subbanda, y los cambios producidos en la RIR no afectan de igual forma a todo el margen frecuencial. En algunos casos, la relación entre la señal a reproducir por los altavoces y la captada por los transductores receptores es no lineal. La solución estándar para este problema de identificación no lineal se basa normalmente en los filtros de Volterra, y esta Tesis Doctoral presenta dos novedosas estrategias de combinación ad-hoc para su utilización en este contexto, las cuales obtienen ventajas de las particularidades de este tipo de filtros. Además, se propone un esquema que presenta una gran robustez con respecto a la ausencia o presencia de distorsión no lineal, e incluso con respecto a variaciones en la potencia de esta distorsión, con un modesto incremento de coste computacional con respecto al de un filtro de Volterra clásico. En muchas ocasiones, la longitud de la RIR es grande y la distribución de su energía no uniforme. Se propone un esquema que, explotando el compromiso entre sesgo y varianza, permite ganancias en esta situación, principalmente cuando la relación señal a ruido es baja. Para mostrar las ventajas del uso de los esquemas de combinación propuestos, se han llevado a cabo una serie de experimentos utilizando un escenario de cancelación de eco acústico monocanal. En todos los casos, las soluciones presentadas han obtenido resultados satisfactorios, demostrando la versatilidad y el potencial de estos algoritmos, y permitiendo mejorar el funcionamiento de los filtros adaptativos ante los condicionantes anteriormente citados. ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Acoustic signal processing applications are becoming increasingly important. Most of these applications, such as acoustic echo cancellation, noise cancellation, dereverberation, separation and tracking of acoustic sources, etc., requires the identification of a (or several) room impulse response (RIR). This response is usually time-varying, what justifies the use of adaptive algorithms to carry out the identification task. The use of adaptive schemes in RIR identification scenarios is subject to different compromises, such as the well-known compromise between speed of convergence and steady-state precision. Several of these tradeoffs are shared by other applications, while others are specific to acoustic signal processing. Among the different methods available to alleviate these limitations, adaptive combination of adaptive filters has been recently receiving a lot of attention, mainly because of its simplicity, versatility, and effectiveness. In this Ph. D. Thesis, we deal with the development, study and implementation of adaptive combination schemes that are especially suited to acoustic signal processing applications. For this purpose, we propose and analyze combination schemes that offer robustness and a suitable behavior with respect to the peculiarities of the involved signals and RIRs. Among all possible determining factors and their potential solutions, in this Ph. D. Thesis we consider: The signal to noise ratio is usually unknown a priori and it can be time-varying. In order to deal with this situation, two new different schemes are proposed. The spectrum of acoustic signals (music and speech) is not flat, what slows down the convergence of adaptive filters. We present a combination algorithm in the frequency domain that allows to mix different frequency bands independently, offering gains that exploit the frequency dependent signal to noise power ratio and the fact that RIR changes can also take place in a frequency-localized manner. Occasionally, the relationship between the signal to be reproduced by the loudspeakers and the signal received by the microphones is nonlinear. The standard solution for this nonlinear identification problem is frequently based on Volterra filters. The Thesis presents two novel ad-hoc combinations strategies to be used in this context, which take advantage of the particularities of this kind of filters. In addition, we propose an additional algorithm that shows great robustness with respect to the presence or absence of nonlinear distortion, and even with respect to changes in the power of nonlinear distortion, with a very modest increment in terms of computational cost. In many cases, very large RIRs are present, and their energies are typically distributed in a non-uniform manner. We propose a scheme that, exploiting the tradeoff between bias and variance, permits important gains in this situation, mainly for low signal to noise power ratios. In order to illustrate the advantages of the proposed combinations schemes, several experiments have been carried out considering a single-channel acoustic echo cancellation scenario. The satisfactory results obtained by the presented solutions demonstrate the versatility and potential of these algorithms, allowing to improve the performance of adaptive filters in the presence of the aforementioned conditions.
Acoustic signal processing applications are becoming increasingly important. Most of these applications, such as acoustic echo cancellation, noise cancellation, dereverberation, separation and tracking of acoustic sources, etc., requires the identification of a (or several) room impulse response (RIR). This response is usually time-varying, what justifies the use of adaptive algorithms to carry out the identification task. The use of adaptive schemes in RIR identification scenarios is subject to different compromises, such as the well-known compromise between speed of convergence and steady-state precision. Several of these tradeoffs are shared by other applications, while others are specific to acoustic signal processing. Among the different methods available to alleviate these limitations, adaptive combination of adaptive filters has been recently receiving a lot of attention, mainly because of its simplicity, versatility, and effectiveness. In this Ph. D. Thesis, we deal with the development, study and implementation of adaptive combination schemes that are especially suited to acoustic signal processing applications. For this purpose, we propose and analyze combination schemes that offer robustness and a suitable behavior with respect to the peculiarities of the involved signals and RIRs. Among all possible determining factors and their potential solutions, in this Ph. D. Thesis we consider: The signal to noise ratio is usually unknown a priori and it can be time-varying. In order to deal with this situation, two new different schemes are proposed. The spectrum of acoustic signals (music and speech) is not flat, what slows down the convergence of adaptive filters. We present a combination algorithm in the frequency domain that allows to mix different frequency bands independently, offering gains that exploit the frequency dependent signal to noise power ratio and the fact that RIR changes can also take place in a frequency-localized manner. Occasionally, the relationship between the signal to be reproduced by the loudspeakers and the signal received by the microphones is nonlinear. The standard solution for this nonlinear identification problem is frequently based on Volterra filters. The Thesis presents two novel ad-hoc combinations strategies to be used in this context, which take advantage of the particularities of this kind of filters. In addition, we propose an additional algorithm that shows great robustness with respect to the presence or absence of nonlinear distortion, and even with respect to changes in the power of nonlinear distortion, with a very modest increment in terms of computational cost. In many cases, very large RIRs are present, and their energies are typically distributed in a non-uniform manner. We propose a scheme that, exploiting the tradeoff between bias and variance, permits important gains in this situation, mainly for low signal to noise power ratios. In order to illustrate the advantages of the proposed combinations schemes, several experiments have been carried out considering a single-channel acoustic echo cancellation scenario. The satisfactory results obtained by the presented solutions demonstrate the versatility and potential of these algorithms, allowing to improve the performance of adaptive filters in the presence of the aforementioned conditions.
Description
Keywords
Procesamiento de señales acústicas, Filtro adaptativo, Algoritmos