Automatic Speaker Verification on Myanmar Spoofing Voice Data using GMM-UBM and TDNN

Artificial voices or human voice imitation pose a risk to speech verification security systems. This study investigates the effectiveness of an automatic speaker verification that utilizes converted voices, aiming to assess the susceptibility of these systems to this type of deception, as voice conversion can be regarded as a variant of voice imitation. Voice conversions among three females are used in the tests. The system performance is assessed with Gaussian Mixture Model-Universal Background Model (GMM-UBM) and Time Delay Neural Network (TDNN) model on Mel Frequency Cepstral Coefficients (MFCC) acoustic features. The objective is to evaluate the resilience of an automatic speaker verification system when subjected to three different voice conversion methods: FreeVC, GMMVC, and Differential GMM (DIFFGMM). The findings indicate that the converted voice has a high rate of error, just as the real voice has fewer error rate.