Artificial neural networks for speaker recognition

As humans, we develop the ability to recognise people by their voice from an early age. Getting computers to perform the same task has proven to be a challenging and interesting problem. Automatic Speaker Recognition is the process of automatically determining who the speaker is (identification) or verifying the claimed identity of a speaker on the basis of information gathered from the spoken words. The advantage of Automatic Speaker Recognition over other biometric techniques is that authentication may be performed remotely. In this thesis, a series of experiments are presented exploring the utilisation of Adaptive Resonance Theory (ART) for the task of speaker identification. The database of speakers includes 12 speakers saying twice the following words: one, two, three, four. They were cut up into frames which contained mel-frequency cepstral coefficients and cepstral derivatives. These cepstral coefficients and/or cepstral derivatives were then presented to an Adaptive Resonance Theory neural network together with an indication of the speaker-id for learning purposes. A number of classifier structures were explored and reported within for the classification of speakers with experimental results confirming the most suitable structure (one speaker - multiple networks). A focus of the thesis is also: (a) in the exploration of fusion techniques for use with multiple classifier structures and, (b) on the vectors used for training and testing, amount and component of the vectors and to see how will they affect the performance of the network. The original objective of the work was to construct neural network(s) that may be employed as speaker identifiers. Whereas this has been achieved the work has shown that great care must be taken to ensure that the size of the input vector used is crucial because it influences the learning process