Model-based methods for linear and non-linear audio signal enhancement

• Main Author: Fong, W. N. W.
• Published: University of Cambridge 2003
• Subjects: 534
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.599095

Owing to the random nature of audio signals, most of the enhancement methodologies reviewed in this work are based explicitly on a Bayesian model-based approach. Of these, the Kalman filter is the most commonly adopted enhancement strategy for a linear and Gaussian restoration problem. To copy with the general non-linear and non-Gaussian case, different filters such as the extended Kalman filter and the Gaussian sum filter have been proposed in the past few decades. As computing power increases, more computationally expensive simulation based approaches such as Monte Carlo methods have been suggested. The main focus of this work is on sequential estimation of the underlying clean signal and system parameters given some noisy observations under the Monte Carlo framework. This class of method is known as sequential Monte Carlo methods, also commonly referred to as the particle filter. In this work, different improvement strategies have been developed and described to improve on the generic particle filtering/smoothing algorithm. A block-based particle smoother is proposed to reduce the memory capacity required for the processing of lengthy datasets, such as audio signals. A Rao-Blackwellised particle smoother is developed to improve on the simulation results by reducing the dimension of the sampling space and thus the estimation variance. To cope with the non-linear restoration problem, a non-linear Rao-Blackwellised particle smoother is developed, which marginalises the parameter state, instead of the signal state as suggested earlier. Finally, we propose an efficient implementation of the suggested slow time-varying model under the sequential Monte Carlo framework for on-line joint signal and parameter estimation.