Adaptive filtering methods for acoustic noise reduction and noisy speech recognition

Abstract

Although adaptive filtering has been successfully used in various fields, it still needs to satisfy more strict requirements to extend its applications. In this dissertation, adaptive filtering was studied to develop several approaches which overcome problems with very long adaptive filter lengths such as unsatisfactory performance, slow convergence, and too much computational demands to achieve real-time processing. The approaches also include methods which surpass conventional approaches to independent component analysis (ICA) and an efficient method which is designed for noisy speech recognition. In order to prove their effectiveness, adaptive noise canceling (ANC) and blind source separation (BSS) are served as benchmarking problems. First, an ANC algorithm based on ICA is derived. Although the most widely used least-mean-square (LMS) algorithm removes noise components based on second-order correlation, there may exist many components which depend on a reference signal through higher order statistics. The ICA-based approach can take these components into account and provided much higher signal-to-noise ratios (SNRs) in the system output than the conventional LMS algorithm. Second, transform-domain adaptive filtering (TDAF) is considered as a method to enhance convergence rates. Stochastic gradient algorithms including the LMS algorithm and the ICA-based algorithm show slow convergence speed especially for colored input signals. TDAF can speed up convergence by pre-whitening input data using unitary transform and improved convergence speed of the ICA-based approach. Third, a uniform filter bank approach is presented. Decimation in a filter bank provides faster convergence rates by making input signals more whitened, and it reduces computational complexity by performing adaptive filtering at a subsampled sampling rate and with much shorter adaptive filters in each subband. As an approach to ICA, the filter bank approach achieved much better performance th...