Mitigating effects of noise in Forensic Speaker Recognition

Among the different biometrics, speech has a significant role in forensics, since it can become a strong evidence in most of the crimes happening today. In Forensic Speaker Recognition (FSR), this speech evidence is used to recognize its speaker. This speech may be unpredictably distorted compared to the speech samples available for other speaker recognition applications. Some of the major distortions affecting the forensic speech quality are noise, speech coding, channel effects, multiple speakers, voice disguise, voice forgery, duration of the speech etc. A forensic speaker verification system which is highly robust to noise is designed and implemented in this paper. Robustness is achieved by employing an iterative threshold based Voice Activity Detector (VAD) in the preprocessing stage and Gammatone Frequency Cepstral Coefficients (GFCC) as the feature set. Gaussian Mixture Model (GMM) which are adapted from a Universal Background Model (UBM) is employed for classification. Performance of the proposed FSR is evaluated under different SNR conditions using equal error rate (EER). Verification accuracy of the proposed system is analysed and compared using Linear Predictive Coding (LPC) coefficients and Mel Frequency Cepstral Coefficients (MFCC) also.