Speaker identification using feature vector reduction of row mean of different transforms

Abstract

In this paper a novel approach to text dependent speaker identification based on feature vector reduction technique of the row mean is proposed. Five different Orthogonal Transform Techniques: Discrete Fourier Transform (DFT), Discrete Cosine Transform (DCT), Discrete Sine Transform (DST), Discrete Hartley Transform (DHT) and Walsh Hadamard Transform (WHT) are applied on the framed speech signal. Feature extraction in the testing and matching phases has been done by using feature vector reduction technique applied on the row mean vector of the magnitude of the transformed speech signal. Two similarity measures Euclidean distance and Manhattan distance are used for feature matching. The results indicate that the accuracy using both the similarity measures remains steady up to certain reduction in feature vector permitting to reduce feature vector size. This algorithm is tested using two databases: a locally created database and CSLU Database. It is observed that, DFT allows maximum percentage of feature vector reduction. It out performs other Transforms with a big margin.