Minimizing the false alarm probability of speaker verification systems for mimicked speech

Speaker verification (SV) systems need to be robust to mimicked voices of target speakers as non-target trials to make them usable in critical applications. However, the performance of SV systems for mimicked voice test conditions has not been extensively explored. In an earlier work, we used Amrita SRE database to evaluate the performance of different state-of-the-art speaker verification systems with mimicked speech: ivector with cosine distance scoring (i-CDS), i-vector with a backend support vector machine classifier (i-SVM) and cosine distance features with SVM backend classifier (CDFSVM), developed using mel frequency cepstral coefficients (MFCC), power normalized cepstral coefficients (PNCC) and delta spectral cepstral coefficients (DSCC). From the experimental results it was observed that the i-SVM system achieved the best overall performance for any input features, MFCC, PNCC or DSCC used. When speaker verification systems are challenged with mimicked voice utterances, the false alarm probability (FAP) of the system increases drastically. In this work, we evaluate the effectiveness of gammatone frequency cepstral coefficients (GFCC) as an input feature in reducing the FAP. Experimental results show that the CDF-SVM with an intersection kernel developed using GFCC gives the best FAP, making the system robust to mimicked speech.