A study on dimensions of feature space for text-independent speaker verification systems

This paper studies the effect of feature dimensions on an MFCC/IMFCC-GMM based text-independent speaker verification system (SVS). A typical baseline system is used to evaluate the impact of features based on the number of Mel, inverted Mel, delta and double delta coefficients while keeping other system parameters constant for all experiments. The relevance of the spectral information contained in the features according to their discrimination power was assessed through a GMM-UBM system with the TIMIT corpus. A new scoring method is reported in which the fusion of feature likelihoods is conducted before the UBM normalization. The study shows that features carrying high frequency spectral content have high information gain enabling better performance of the SVS. Similarly, adding more coefficients of MFCCs and IMFCCs instead of dynamic features such as delta and double delta coefficients improves the SVS´s equal error rate (EER). Our scoring technique outperformed the traditional scoring algorithm by 9.7%.