Adaptive filtering for noisy and mismatched conditions in speaker recognition

Abstract

This paper presents comparison of different adaptive filter algorithms i.e. Wiener, Least Mean Square LMS, Normalized Least Mean Square NLMS, Block Least Mean Square, BLMS, as system identifiers of telephonic channel; specially when there is a mismatch in train and test conditions of speaker recognition system. We need to know the response of this telephonic channel in order to minimize the mismatch between training and testing data. Once this response is known we can get filter weights either to model telephonic channel and pass the clean speech through it or make and inverse system out of the identified coefficients in order to enhance the corrupted signal. Results are obtained on white Gaussian noise added signal and signal passed through telephonic channel. It has been seen that LMS and NLMS perform better over white Gaussian noise added signals whereas in case of mismatched train/test conditions, the better estimate of unknown telephonic channel response was given by LMS algorithm with small step size of 0.003 and then at 0.5, 0.6 step size. Effect of varying step size has been observed not only for LMS but it's variants as well. Overall results show significant improvement from 20% to 80-66% in speaker recognition accuracy.