Comparison of LPCC and MFCC features and GMM and GMM-UBM modeling for limited data speaker verification

Abstract

This work address text-independent speaker verification with the constraint of limited data (<;15 seconds). The existing techniques for speaker verification work well for sufficient data (>1 minute). Developing techniques for verifying the speakers for limited data condition is a challenging issue since data available of speakers is very small nowadays. This is because people reluctant to give more data. In this work to extract features of speech signal Mel-Frequency Cepstral Coefficients (MFCC) and Linear Prediction Cepstral Coefficients (LPCC) are used. The extracted features are modeled using Gaussian Mixture Model (GMM) and GMM-Universal Background Model (UBM) modeling techniques. The NIST-2003 database is used to carry-out the experiments. The experiments are evaluated for limited amount of training and testing speech data. The experimental observation indicates that the Equal Error Rate of LPCC features is less as compared to MFCC for limited data.