Fast pole filtering for speaker recognition

Mismatched training and testing conditions for speaker recognition exist when speech is subjected to a different channel for both cases. This results in diminished speaker recognition performance. Estimating and removing the channel filtering effect will make speaker recognition systems more robust. It has been shown that a reliable estimate is obtained by taking the mean of the pole filtered linear predictive (LP) cepstrum. Finding the pole filtered mean requires factorization of the LP polynomial which is computationally intensive especially for real time applications. In this paper, we examine a fast method of doing pole filtering that avoids polynomial factorization. This method is much more computationally efficient and gives equal or better performance than the conventional way of doing pole filtering. Experimental results are given for four databases having a variety of mismatched conditions.