Adapting to Noise in Forensic Speaker Verification Using GMM-UBM I-Vector Method in High-Noise Backgrounds

The performance of the GMM-UBM-I vector in a forensic speaker verification system has been examined in the context of noisy speech samples. This analysis utilised both Mel-frequency cepstral coefficients (MFCC) and MFCCs generated from auto-correlated speech signals. The noisy signal’s auto correlation coefficients are concentrated around the lower lag, whereas the autocorrelation coefficients near the higher lag are very small. Thus, in addition to retain the periodic nature, autocorrelation-based MFCC is also robust for analyzing speech signals in intense background noise. The performance of MFCC and auto-correlated MFCC depends heavily on the quality of the sample. It works best with data that is free of noise, but it suffers when used on real-world examples, ie, with noisy data. The experiment on speaker verification for forensic purposes involved the addition of White Gaussian Noise, Red Noise, and Pink Noise, with a Signal-to-Noise Ratio (SNR) range spanning from −20 dB to + 20 dB. The performance of both methods was affected drastically in call cases but autocorrelation-based MFCC gave better results than MFCC. Thus, autocorrelation-based MFCC is a valuable method for robust feature extraction when compared with MFCC for speaker verification purposes in intense background noise. The verification accuracy in our method is improved even in very high noise levels (−20 dB) than the reported research work.