Speaker recognition with a neural classifier

Describes a new approach to speaker recognition where a neural classifier is used to separate the different speakers. This has the advantage of being able to incorporate into an individual model more information about the speaker population generally, thus allowing the training process to discard classes of speech that are common to many speakers and focus on those classes that are the best discriminators. Results show that the neural classifier achieves at least comparable performance to a conventional approach based on vector quantisation and codebooks. Detailed comparisons in terms of speaker identification error rates, levels of training, and text dependency are presented