Information fusion for robust speaker verification

In this paper we have studied two information fusion approaches, namely feature vector concatenation and decision fusion, for the task of reducing error rates in a speaker verification system used in mismatched conditions. Three types of features are fused: Mel Frequency Cepstral Coefficients (MFCC), MFCC with Cepstral Mean Subtraction (CMS) and Maximum Auto-Correlation Values (MACV). We have used the mismatch sensitivity of Linear Prediction Cepstral Coefficients (LPCC) as a speech quality measure for selecting the weight of the contribution of the MFCC modality in the adaptive decision fusion approach. We show that in most cases concatenation fusion is superior to decision fusion. The results lead us to propose a hybrid fusion approach in which two combinations of concatenation fusion are further fused using adaptive decision fusion. The hybrid system is shown to have the lowest error rates on both clean and noisy speech.