Text Dependent Speaker Recognition Using Linear Prediction Coefficients-Dynamic Time Warping

Abstract

This project proposes a text-dependent speaker identification system. Isolated digits 0-9 and their concatenations are used for speaking text. For each speech signal Linear Prediction Coefficients (LPC) are extracted and formed as feature vectors. Dynamic Time Warping (DTW) is used to measure distances between referenced and evaluated vectors. These distances, indicating nearness of unknown vectors to references, incorporated with K-Nearest Neighbor (K") decision technique are used for the identification process. In the verification test of the experiment. Consequently, we have experimented on the use of LPC with both DTW (using KNN as a decision rule) and ANN (a well-known Multilayer Perception (MLP) with back propagation learning algorithm). The systems were tested with 0-9 isolated digits. It has been shown that DTW with KNN gives better performance. It is affected by an attempt of ANN to recognize all of training patterns including any low quality voice. In this paper, further successive progress of our system is to deeply experiment on the use of DTW with KNN with concatenated digit, which will be a form of speaking-text in our application at last. This research will also purpose in selection of some acceptable digits to be included in our text-prompted speaker identification system.