Robust speaker recognition for e-commerce system

In banking system of e-commerce, where privacy and authentication are more important and voice based authentication is preferred, telecommunication plays vital role by enabling users to interact remotely and verify their identities. The design of such robust speaker recognition system can have various problems related to combined optimization of a number of processes like front-end noise removal, auditory parameter extraction and the speech modeling/recognition paradigm. In this paper we examine these front-end noise reduction/removal techniques with focus on development of a generic speech improvement method based on hybrid echo filter to particularly handle speech signals that travel through telephonic channel. Firstly noised telephonic signals are subjected to spectral subtraction and accuracy of correctly denoised samples is obtained which is 76%. The resultant signals are further improved by applying echo filter technique which dramatically increases the accuracy up to 83.5%. Achieved results are highly significant as they improve accuracy of the system as compared to contemporary work in the area.