Maximum-likelihood stochastic matching approach to non-linear equalization for robust speech recognition

We present a new technique in the stochastic matching framework to compensate for nonlinear distortions in speech recognition. The features of the test data and the means of the trained model are both transformed using neural networks to better fit each other. The parameters of the neural network are estimated using a novel combination of the generalized EM (GEM) and the backpropagation algorithms. In the feature transformation case, when the exact Q-functions cannot be calculated, approximations are heuristically derived. The mathematical properties of the new algorithm are analysed. The performance of the algorithm is also studied under different mismatch conditions