ART Network Based Text Independent Speaker Recognition System for Dynamically Growing Speech Database

摘要

Automated recognizing a speaker from the speech signals is the foremost application in forensics. Speaker recognition system involves two phases namely feature extraction and a classifier system. Features extracted from the speech signals are fed to an already trained classifier system that identifies the speaker. Major challenge occurs when the database is periodically updated which necessitates retaining the classifier with new set of exemplars includes the old and new datasets. As training the neural network is computationally intensive, Back Propagation system is not ideal for speaker recognition system (updation). Hence it necessitates an efficient speaker recognition system that doesn't forget the old database but adjusts to the new set of data. In this paper an Adaptive Resonance Theory (ART) based speaker recognition system is proposed that is capable of functioning well even in the case of periodic updation.