Robust speech recognition by improvement missing features using Bidirectional Neural Network

In this paper we present a new method for nonlinear compensation of mismatches, e.g. additive noise, on clean and noisy speech recognition. We were inspired by the human recognition system in development and implementation of a new Bidirectional Neural Network (BNN). This procedure, results in improvement of input features and consequently increasing the overall recognition accuracy. The feedforward weights of this network are trained using both clean and noisy speech features. The results demonstrate significant improvements in clean and especially noisy speech recognition accuracy compared to reference model trained on unimproved features.