Abstract
In this thesis, a noise reduction algorithms is presented from the perspective of source localization and separation. Minimum Power Distortionless Response (MPDR) algorithm is utilized to determine the bearings of the signal and noise sources. Tikhonov regularization (TIKR) and compressive sensing (CS) algorithm are employed to extract the amplitudes of the signal and noise sources. In order to evaluate the proposed method, the log minimum mean-square error (log-MMSE) algorithm, the Generalized Sidelobe Canceller (GSC), and the regulated multiple-input/output inverse theorem (R-MINT) are adopted as benchmarking methods. The Log-MMSE is used to estimate an optimize gain correction function as a post-filter. To enhance GSC, subband (SB) filtering and internal iteration (IIT) are incorporated, which is termed the GSC-SB-IIT method. The R-MINT used to be applied in room response inverse filtering. Numerical simulations and experiments are conducted for a 24-channel uniform circular microphone array. White noise and traffic noise are used in simulating the background noise. Objective tests based on the segmental signal-to-noise ratio (segSNR) and Perceptual Evaluation of Speech Quality (PESQ) and subjective listening tests are conducted to compare the noise reduction approaches. The results show that the CS algorithm has achieved the highest reduction of noise.