SPEECH RECOGNITION USING NEURAL NETWORKS

Abstract

Although speech recognition products are already available in the market at present, their development is mainly based on statistical techniques which work under very specific assumptions. The work presented in this thesis investigates the feasibility of alternative approaches for solving the problem more efficiently. A speech recognizer system comprised of two distinct blocks, a Feature Extractor and a Recognizer, is presented. The Feature Extractor block uses a standard LPC Cepstrum coder, which translates the incoming speech into a trajectory in the LPC Cepstrum feature space, followed by a Self Organizing Map, which tailors the outcome of the coder in order to produce optimal trajectory representations of words in reduced dimension feature spaces. Designs of the Recognizer blocks based on three different approaches are compared. The performance of Templates, MultiLayer Perceptrons, and Recurrent Neural Networks based recognizers is tested on a small isolated speaker dependent word recognition problem. Experimental results indicate that trajectories on such reduced dimension spaces can provide reliable representations of spoken words, while reducing the training complexity and the operation of the Recognizer. The comparison between the different approaches to the design of the Recognizers conducted here gives a better understanding of the problem and its possible solutions. A new learning procedure that optimizes the usage of the training set is also presented. Optimal tailoring of trajectories, new