Limited data speaker verification using MFSR analysis technique

Abstract

In this paper, the task of verifying the speaker using limited training and testing data is addressed. Speaker verification under limited data may not be able to produce sufficient feature vector for training and testing. This creates poor speaker modelling during training and testing. To defeat this problem, feature vectors for training and testing are increased. To increase the feature vectors, multiple frame size (MFS), multiple frame rate (MFR) and multiple frame size and rate (MFSR) analysis techniques are explored. These techniques comparatively increase more feature vectors during training and testing compared to single frame size and rate (SFSR) analysis. With the help of these feature vectors improved modeling and testing can be done under limited data. To demonstrate this we have used Mel-frequency cepstral coefficients (MFCC) as feature extraction technique. Gaussian mixture modelling-Universal background model (GMM-UBM) is used for modelling the speaker. The NIST-2003 is used as database for conducting the experiments. The experimental results show that there is an improvement in the performance of speaker verification if we use MFS, MFR and MFSR are analysis techniques instead of SFSR.Further, the experimental results show that MFSR gives improved performance over other analysis techniques.