A Comparative Study of Text-Independent Speaker Identification using Statistical Features

This paper concerns a comparative study on long term text-independent speaker identification using statistical features. Performances of six statistical methods are compared. Four of them are the distance measures (the City block, the Euclidean, the Weighted Euclidean, and the Mahalanobis distance measures). The other two are the Gaussian probability density estimation and the probability estimation after the Karhunen-Loeve orthogonal transformation. Comparisons are based on two statistical tests (the Friedman test and the multiple comparison approach). Experimental results show that (1) probability calculation is generally better than most distance measures, (2) the orthogonally transformed feature vectors of dimension 15 (originally 20) performs better than all the other methods, (3) the Weighted Euclidean distance measure performs file:///C|/Documents%20and%20Settings/Ponn/Desktop/ijcim/past_editions/1998V06N1/article5.htm (1 of 12)24/8/2549 12:08:59 A Comparative Study of Text-Independent Speaker Identification using Statistical Features better than the other three distance measures, and (4) the Mahalanobis distance measure does not perform well. An explanation is advanced for this result in the conclusion.