Spectral Features based Robust Speech Spoofing Detection System
Abhinav Sharma, Anshu Sharma, Taruna Anand, Pradeep Kumar Juneja, Anushka Agarwal · 2022
Speech Replay can be used, with the recorded speech of any individual, for spoofing or gaining fraudulent access through an automatic speaker verification system. In this research work, Speech spoofing detection has been developed in MATLAB platform that detects the spoofing by identifying between two classes a speech may belong to, namely the Original (genuine) and the Recorded then Replayed (R-R) Speech. This work develops and implements the detection of the speech spoofing, carried out in English language with the use of spectral features (Mel Frequency Cepstral Coefficients-MFCC's) of the speech samples recorded, in two classes, as mentioned above (O, R-R). Classification of speech samples is performed by Gaussian Mixture Models (GMM). In the proposed work, the original (noiseless) speech samples and their R-R versions were used to train the GMM's for 2 classes and the testing work has been performed using a large corpus of (i) original (almost noiseless), (ii) the concatenated versions of originals of both classes with various noises such as crowd-talking, fan, traffic, light, medium and heavy rain, construction site etc. and also (iii) the random samples collected at various public places, to test the robustness of the system. Optimum detection is observed by comparing the performance of the GMM with variation in (i) the number of MFCC, (ii) Gausses (components), (iii) the elapsed time and (iv) number of iterations. The average spoofing detection performance is resulted to be 95.93 percent.