Combining independent component analysis with geometric information and its application to speech processing

In this paper, we propose two approaches for combining geometric information with ICA algorithm to solve permutation problem under the scenario where a rough information about the direction of the desired source is known. The first approach is a new blind extraction algorithm with a soft quadratic geometric constraint. The desired source is guaranteed to be conveyed to the output with little distortion by the quadratic constraint and the negentropy maximization criterion is used to ensure that the other sources get suppressed at the output. The second approach employs a quadratic geometric test as a post-processing step to pickup the desired source after ICA processing. An advantage of the proposed two approaches is that they do not require accurate knowledge of the number of sources in the mixtures to recover the desired source, in contrast, other geometric ICA approaches usually fail if the number of sources is not known accurately.