Signal Separation Besed on Independent Component Analysis with Layered Neural Network and Its Adaptability to Untrained Vocal Signals

Abstract

Independent component analysis (ICA) is a signal separation technique inspired by the famous psychological phenomenon called cocktail party effect. Various kinds of its applications have been undertaken by a lot of researchers so far, and an alternative method based on a layered neural network with structural pruning was tried in the preceding studies. However, how to develop such a signal separation matrix was the center of attention, so how to apply it after training was not discussed a lot. Then, from the viewpoint of adaptability to untrained signals, some computer simulations are carried out in this study. As a result, it is found experimentaly that a vocal signal separation task with the developed separation matrix is accomplished successfully as we have intended in advance.