Fast adaptive component weighted cepstrum pole filtering for speaker identification

Mismatched training and testing conditions for speaker identification exist when speech is subjected to a different channel for the two cases. This results in diminished speaker identification performance. Finding features that show little variability to the filtering effect of different channels will make speaker identification systems more robust thereby achieving a better performance. It has been shown that subtracting the mean of the pole filtered linear predictive (LP) cepstrum from the actual LP cepstrum results in a robust feature. This feature is known as the pole filtered mean removed LP cepstrum. Another robust feature is the adaptive component weighted (ACW) cepstrum particularly with mean removal. In this paper, we combine the ACW cepstrum with the pole filtering concept to configure a more robust new feature, namely, the pole filtered mean removed ACW cepstrum. This new method is fast and shows a higher performance then the pole filtered mean removed LP cepstrum and the mean removed ACW cepstrum. Experimental results are given for the TIMIT database involving a variety of mismatched conditions.