A speaker verification system based on a neural prediction model

Describes the use of a neural prediction model to perform the task of speaker verification. The model uses a sequence of multi-layer perceptrons as a separate nonlinear predictor for each speaker known to the system. It is designed to represent temporal structures of speech patterns as cues in speaker verification. Temporal distortion of speech is efficiently normalised by a dynamic programming technique. Speaker verification and model training algorithm are presented based on a combination of dynamic programming and back propagation techniques. Experiments are carried out to investigate the performance of the neural prediction model as a speaker verification system and to investigate its vulnerability to changes in speech parameters over time. A speaker verification accuracy of 100% is achieved, albeit with a relatively small test set of 60 utterances from one true speaker and five impostors. A comparison with other methods, such as vector quantisation and MLP discriminator is also presented