VOICE RECOGNITION USING ARTIFICIAL NEURAL NETWORKS AND GAUSSIAN MIXTURE MODELS

摘要

The ability of recognition systems to correctly recognize speakers based on their speech waveform distribution depends largely on how the recognition system can train the model parameters so as to provide the best class of discrimination. This paper presents the results of an effort to recognize the voice of individual speakers based on their continuous speech waveform distribution using the combined frameworks of artificial neural networks (ANN) and statistical Gaussian mixture models (GMM). A feed-forward multilayer ANN architecture with 30 hidden neurons was implemented for discriminative classification and training and the statistical GMM model computed scores that were transferred to best match the speech features. The decision system determines the recognized speakers using correlation coefficient analysis to measure the goodness of match of speech feature frames of the detected speaker from the ANN and GMM frameworks. To validate performance of the system, experiments were conducted using speech utterances from 30 different speakers (20 males and 10 females). System performance showed average recognition rates of 77% for 5-word utterances and 43% when the lengths of the utterances were increased to 20-word utterances for cases of trained speech utterances. With unknown utterances, recognition rate of 18% achieved for 20-word utterances.