Evaluation of Blind Separated Signals Using Speech Recognition System

Output blind separated signals are recovered independent sources obtained from their mixtures using methods of convolutive independent component analysis. These methods have no information about the original sources and the mixing process. For the separation process we adapted and improved a method which is based on the natural gradient approach. The new concept proposed in this paper then tries to measure the hit rate of the source restoration by means of the discrete word intelligibility using an Internet speech recognition system. We are interested in a two-channel case, in which we observe the outputs of a 2 times 2 linear time-invariant system of mixtures of speech, music and noise signals