Phoneme Based Speaker Verification System Based on Two Stage Self-Organizing Map Design

Speaker verification is one of the pattern recognition task that authenticate a person by his or her voice. This thesis deals with a relatively new technique of classification that is the self-organizing map (SOM). Self-organizing map, as an unsupervised learning artificial neural network, rarely used as final classification step in pattern recognition task due to its relatively low accuracy. A two-stage self-organizing map design has been implemented in this thesis and showed improved results over conventional single stage design. For speech features extraction, this thesis does not introduce any new technique. A well study method that is the linear prediction analysis (LP A) has been used. Linear predictive analysis derived coefficients are extracted from segmented raw speech signal to train and test the front stage self-organizing map. Unlike other multistage or hierarchical self-organizing map designs, this thesis utilized residual vectors generated from front stage self-organizing map to train and test the second stage selforganizing map. The results showed that by breaking the classification tasks into two level or more detail resolution, an improvement of more than 5% can be obtained. Moreover, the computation time is also reduced greatly