The fundamental frequency (F0) or pitch can be extracted from only the voiced part of the time unit in the utterance. Therefore, the F0 needs to be interpolated in unvoiced regions to avoid variance problems in recognition using a smoothed log-pitch estimate and its two temporal derivatives [2]. In this paper, the average magnitude difference function (AMDF) is used instead of autocorrelation function to extract the pitch period, as used in [15]. It computes the difference between the signal and time shifted version of itself. The average magnitude difference function [16] is defined as: