Robust speech recognition by modifying clean and telephone feature vectors using bidirectional neural network

In this paper we present a new method for nonlinear compensation of distortions, e.g. channel effects and additive noise, in clean and telephone speech recognition. A Bidirectional Neural Network (BidiNN) was developed and implemented in order to modify distorted input feature vectors and improve the overall recognition accuracy. Distorted components in feature vectors were estimated in accordance with the latent knowledge in the hidden layer of the neural network. This knowledge is obtained by training with clean and telephone speech, simultaneously and is mostly induced by phonemic content and less influenced by the irrelevant variations in speech signal. An MLP neural network was trained with these modified feature vectors. Comparing the achieved results with a reference model that was trained with unmodified feature vectors, demonstrate significant improvement in clean and telephone speech recognition accuracy.