Automatic Speaker Verification using Gammatone Frequency Cepstral Coefficients
Himanshu Choudhary, Debanjan Sadhya, Vinal Patel
Abstract
Biometric based technology is often used for identification in access control systems. One of the most important behavioral biometrics methods is speech verification, wherein audio cues are utilized for decision making. Since there is an increase in audio verification systems nowadays, several attackers have developed novel spoofing techniques to capture the data. To resolve this problem, an automatic speaker verification system (ASV) is used. The ASV system authenticates speakers by analyzing speech utterances. This paper presents a probabilistic verification model built on machine learning techniques for accurately detecting and reporting such attacks. We have utilized Gammatone frequency cepstral coefficients (GFCCs) as a prominent feature along-with the existing features like Mel frequency cepstral coefficients (MFCCs) and pitch for improving the ASV system accuracy. These features are subsequently passed through the Gaussian mixture model (GMM) and K nearest neighbor (KNN) classifiers for making the final decision. Among all the results, the lowest EER of 0.07 is obtained when GFCC is used with GMM as the baseline classifier.