| There is an increasing interest in the field of speech recognition because of expanding use of multimedia applications. A number of speech recognition algorithms are available in the literature. The objective of this dissertation is to develop a speech recognition algorithm which can recognize various phonemes. This objective has been accomplished. The proposed algorithm has three stages; adaptive filtering, wavelet transform, and neural network learning. In the adaptive filter stage, the least mean square algorithm is modified in order to find the optimal solution for adaptive filter coefficients from input speech. In the second stage, the input speech signal can be represented in terms of a wavelet expansion. It uses a combination of the coefficients of the wavelet function, and data operations can be performed using the wavelet transform. In the third stage, the neural network improves the intelligence of systems working in an uncertain environment. Neural network consists of two phases, training and recognition. Back-propagation training is utilized in this work. The learning algorithms are used for adjusting coefficients and parameters to approximate desired sets of inputs. For the neural network to find the vector parameters, a learning algorithm was developed based on the delta learning rule and the modified conjugate gradient methods.; The software implementation of the algorithm has been made using Microsoft Visual C. The algorithms are tested by making a database of various phonemes spoken by different individuals. The proposed algorithm is suitable for use under limited noise conditions. |
