Noise immunization using neural net for speech recognition

The multilayer perceptron (MLP) type of neural network classifiers using backpropagation has become increasingly popular for speech recognition. However, for the case of noisy speech, studies have not been very extensive. In this paper, a robust speech recognition system using a neural network is studied. Robustness is achieved by noise immunization, thereby enabling the system to maintain a high recognition accuracy for speech input at different signal-to-noise ratio (SNR) conditions. Noise immunization is achieved by gradual contamination of the signal with noise thereby creating a more reliable reference database in spite of low SNR. The learning is done by a modified backpropagation algorithm. Tenth order LPC coefficients are used to represent the data. The order or sequence in which the data is presented to the neural network for training to provide fast convergence and better performance is studied.<>