Speaker identification by K-nearest neighbors: Application of PCA and LDA prior to KNN

This article presents the task of speaker identification in a closed group. It discusses main steps of the identification process ranging from the proper speech features to the classification methods and statistical signal processing. However, its main focus is on tuning the final system using KNN classification method by setting up the number of neighbors, and reducing the feature vector dimension by PCA and LDA not only to speed up but possibly improve the overall performance. By selecting eligible number of neighbors a 6% improvement in the recognition was reached. Moreover, application of both PCA and LDA reduced the feature vector dimension by more than 50% while slightly increasing the recognition accuracy.