Continuous speech recognition

Abstract

Continuous speech recognition is one of the greatest challenges in this beginning of the centurie. The impressive advances in hardware allow the use of sophisticated mathematical methods to solve complex problems. In this thesis, we show that methods invented for solving long constraint length convolutional codes can be used in speech recognition. The main contributions of this thesis can be summarized as follows: (1) The applicability of the stack decoding algorithm to continuous speech recognition. (2) The development and the analysis of a path metric based on the Mahalanobis distance. This path metric has been used in the implementation of the stack algorithm in recognition program and also in the Viterbi algorithm in the training program. (3) The development of a novel algorithm for clipped speech restoration based on linear prediction (4) Speech denoising method based on time varying Wiener filters. We obtained remarkable results using a very low order filter. (5) In order to develop a good path metric for our stack algorithm, we have performed a statistical analysis of three parametric representations of speech and we have shown that the MFCC set is nearly Gaussian and provides the best separability between classes as compared with LPC and PARCOR coefficients. (6) In the last chapter of the thesis, we have developed an automatic segmentation algorithm that we use for training the speech recognition program