This thesis proposes algorithms to improve the performance of two-channel speech enhancement system. A frequency-domain generalized sidelobe canceller (GSC) and a minimum mean-square error log-spectral amplitude (MMSE-LSA) estimator are mainly conside...
This thesis proposes algorithms to improve the performance of two-channel speech enhancement system. A frequency-domain generalized sidelobe canceller (GSC) and a minimum mean-square error log-spectral amplitude (MMSE-LSA) estimator are mainly considered to implement the two-channel speech enhancement system. We propose a soft-decision adaptation mode controller (SD-AMC) which improves the performance of frequency-domain GSC and a two-channel minimum mean-square error (MMSE) estimator as an efficient structure for two-channel speech enhancement system.GSC is an adaptive algorithm to implement the microphone array system and it needs blocks to decide whether the adaptive filters should be updated or not. A SD-AMC softly and frequency independently controls the adaptation mode of the adaptive filters in frequency-domain GSC and it improves accuracy and robustness of the system.The two-channel speech enhancement system generally combines microphone array and single-channel speech enhancement blocks to guarantee the efficient noise reduction ability. Two-channel MMSE estimator is a method to combine the microphone array algorithms to MMSE-LSA estimator, efficiently. By using input signals of two sensors, the proposed structure estimates the signal-to-ratio and speech presence probability which are key parameters of the MMSE-LSA algorithm efficiently and the well-estimated parameters improve the performance of the MMSE-LSA system.To evaluate the performance of proposed algorithm, speech recognition rate is tested in car environment. The GSC with SD-AMC improves the recognition rate by 14.54 % and the two-channel MMSE structure shows improvement by 17.28 % comparing with the GSC using conventional AMC in speech recognition test.