Text-Dependent Speaker Recognition with long-term features based on functional data analysis

Abstract

Text-Dependent Speaker Recognition (TDSR) is widely used nowadays. The short-term features like Mel-Frequency Cepstral Coefficient (MFCC) have been the dominant features used in traditional Dynamic Time Warping (DTW) based TDSR systems. The short-term features capture better local portion of the significant temporal dynamics but worse in overall sentence statistical characteristics. Functional Data Analysis (FDA) has been proven to show significant advantage in exploring the statistic information of data, so in this paper, a long-term feature extraction based on MFCC and FDA theory is proposed, where the extraction procedure consists of the following steps: Firstly, the FDA theory is applied after the MFCC feature extraction; Secondly, for the purpose of compressing the redundant data information, new feature based on the Functional Principle Component Analysis (FPCA) is generated; Thirdly, the distance between train features and test features is calculated for the use of the recognition procedure. Compared with the existing MFCC plus DTW method, experimental results show that the new features extracted with the proposed method plus the cosine similarity measure demonstrates better performance.