Recently, blind signal separation by Independent Component Analysis(ICA) has been received attention because of its potential applications in various fields of signal processing. ICA finds a linear coordinate system(the unmixing system) so that the resulting signals are statistically independent from each other as possible.
ICA depends on several assumptions such that sources are mutually independent and linearly mixed. These assumptions restrict the performance of ICA within few applications of real environments where all of the assumptions are satisfied. For most applications of ICA in real environments, it is necessary to release of all assumptions or compensate for these restrictions.
In this work, an algorithm to improve the performance of ICA was proposed. It can compensate for the restrictions of ICA by feedback operations from a classifier to the ICA network. Under the assumption of the stationary environment, providing additional information of the mixing environment for the ICA network can improve the recognition performance. The proposed algorithm performs iterative feedback operations to provide the additional information of the classifier for the ICA network.
The algorithm was applied to isolated-word speech recognition in noisy environments. Mel-Frequency Cepstral Coefficient(MFCC) was used for feature extraction and Multi-Layer Perceptron(MLP) was the classifier. For the convolved mixture of speech and noise recorded in real environment, the algorithm improved the recognition performance and showed robustness against parametric changes.