Accuracy Comparison using Different Modeling Techniques under Limited Speech Data of Speaker Recognition Systems

Abstract

Pointing towards programmed machine learning by human, a technique for speaker recognition with speaker identity in light of man machine interface is an interest of science. Motivated by the same, we propose a philosophy to recognize speakers. Inside of our investigation, obtaining speech signal, analysis of spectrogram, neutralization, extraction of speaker specific features for recognition, mapping of speech using Novel Vector Quantization (NFVQ) is presented. NFVQ is particularly suitable for colossal arrangement of information and yield discourse mapping. Furthermore Speaker Recognition by utilizing NFVQ Model additionally will be exhibited in this paper. During feature extraction, traditional triangular shaped bins have been replaced by Gaussian shaped filter (GF) and Tukey filter (TF) to calculate Mel Frequency Cepstral Coefficients (MFCC). This work performs an experimental evaluation of three simple modelling techniques namely, Fuzzy c-means, FVQ2 and NFVQ. Among these NFVQ shows significant improved performance compared to Fuzzy c-means and FVQ2. For about 10 s of training and testing speech data of speakers the efficiency for NFVQ, FVQ2 and Fuzzy c-means are 98.8%, 73.33, and 8, respectively, for a set of 630 speakers taken from the TIMIT database. We additionally got 5% outright EER change for both-sex trials on the 10 s-10 s condition contrasted with the FVQ2 approach.