Speaker Verification Using Cosine Distance Scoring with i-vector Approach

In this paper, a robust yet simple speaker verification system is implemented. The speaker verification system is investigated employing the i-vector approach with the Cosine Distance Scoring (CDS) for system classification. In addition, to measure the system performance, Equal Error Rate (EER), Detection Error Trade-off (DET) Curve, Receiver Operating Characteristic (ROC) curve as well as Detection Cost Function (DCF) were utilized. Experimental results are conducted on the TMIT database using 64 randomly selected speakers. The proposed system utilizes the Mel Frequency Cepstral Coefficients (MFCC) and Power Normalized Cepstral Coefficients (PNCC) for feature extraction. In addition, features normalization methods such as Feature Warping (FW) and Cepstral Mean-Variance Normalization (CMVN) are used in order to mitigate channel effect noise. The speakers are modeled with the i-vector while CDS is used for classification. Experimental results demonstrate that the proposed system achieved promising results while being computationally efficient.