This thesis presents advanced method for adaptive noise canceling(ANC) for speech signals, which is based on independent component analysis(ICA). ANC is an approach to reduce noise in a corrupted signal based on the reference input. The most popular training method for ANC is the least-mean-squares(LMS) algorithm, which minimizes second-order correlation between the corrupted signal and the reference input. However, ANC based on ICA is recently proposed, which shows better performance than the conventional LMS algorithm. In order to reduce noise, this method utilizes ICA which finds independent signals from given mixed signals by considering higher-order statistics. In contrast to the LMS algorithm, ANC based on ICA removes not only second-order correlation but also higher-order dependency between the corrupted signal and the reference input.
In order to apply ANC to speech enhancement, we should assume a special situation and consider some problems. The situation is that the reference microphone is far enough from speech, that is, reference input does not contain the speech signal. The problems are that the channel from noise source to reference input is nonminimum-phase and more than one noise sources exist. To deal with these problems, ICA-based algorithms for noncausal adaptive FIR and IIR filter are proposed and it is extened to the case of more than one reference inputs. Simulation results show that the proposed algorithms are adequate for the considered situations.