Speech Recognition with Frequency Domain LineaR Prediction

This paper deals with the main objective to design an efficient algorithm to enhance speech recognition. The conventional speech recognition does not work well in conditions of noise and reverberation. The accuracy and performance of an automatic speech recognition system is dependent on the feature vector generated which characterizes the signal. The feature extracted must be independent of speaker variability, noise and other parameters. Hence, the only solution is to have a more robust feature vector. The feature vector must be such that it models only the intelligible data and not influenced by artifacts in any form. The proposed feature vector with maximum accuracy is a combination of MFCC (Mel-Frequency Cepstral Coefficients) and FDLP (Frequency Domain Linear Prediction) after applying Short Time Energy (STE) to find the highest energy indexed block of 400 samples. The chosen block is used for analysis. We have used Artificial Neural Network (ANN) for classification of samples (digits from 0-9).