An Approach to Noise Robust Speech Recognition using LPC-Cepstral Coefficient and MLP based Artificial Neural Network with respect to Assamese and Bodo Language

In this paper, a new simplified approach has been made for the design and implementation of a noise robust speech recognition using Multilayer Perceptron (MLP) based Artificial Neural Network and LPC-Cepstral Coefficient. Cepstral matrices obtained via Linear Prediction Coefficient are chosen as the eligible features. Here, MLP neural network based transformation method is studied for environmental mismatch compensation. MLP based neural network has been used by many researchers in conjunction with speech recognition, basically for the transformation of the speech feature vectors. In our current study, neural network (MLP) is used to compensate for the environmental mismatch either in feature domain, the model domain, or both. It has been observed that environmental mismatch is automatically compensated without particular knowledge of the environmental interference and retraining. This method can be applied to both linear and non-linear distortion of the speech signal, such as in noisy reverberant speech or telephone speech. Further it can be used for speaker adaptation. By using MLP based neural network, the adaptation processes would require small volume of training data. The Assamese and Bodo are two local languages of North-East India, and they are used as reference languages to carry out this study. General Terms Speech Recognition, Language, Computing machine