Speaker recognition using parallel neural network modules

This dissertation presents methods for improving the performance of Artificial Neural Network Speaker Recognition systems based on a novel negative reinforcement training algorithm and selective filtering of training data. A brief background of current speaker recognition technology is presented including various options for feature extraction and feature mapping methods. We then develop and present MATLAB implementations of four separate closed-set speaker recognition systems using wavelet and cepstral vector feature extraction algorithms and using several parallel-architecture neural network mapping algorithms. These systems are used as performance baselines representing the most promising feature extraction and feature mapping algorithms currently in use. The negative reinforcement training algorithm and training data selection concepts are then described and demonstrated as improvements to the baseline systems. The resulting systems are shown to have significantly improved correct recognition rate over the baseline systems.