| ¡@¡@In this talk, we consider the overcomplete independent component analysis problem, and assume there is no additive noise in our model. Even though Lewicki and Sejnowski (2000) gave an unsupervised learning algorithm for this problem. But there still exist some computational difficulties in their algorithm. In order to avoid those difficulties, we propose a new algorithm for learning the bases vectors for the observations, and recovering the original sources from the sparse prior assumption. Here we demonstrate that this algorithm works for the blind separation of the speech data. The case of our experiment involve two sequences of observations mixed by the three independent sources. |