Teknik-teknik mengenalpasti sela masa senyap dalam sistem pengecaman suara

Abstract

Classification of speech into voiced, unvoiced and silence (V/UV/S) regions is an important process in many speech processing applications such as speech synthesis, segmentation and speech recognition system. Two such measures are investigated with respect to their ability to discern voiced/unvoiced and silence segments of speech. They are the Instantaneous Energy (IE) and Local Time Correlation (LTC) method. Both IE and LTC methods are recently proposed technique for nonstationary signal analysis and have been successfully applied to speech processing. A comparative study was made using these two algorithms for classifying a given speech segment into two classes: voiced/unvoiced speech and silence. IE and LTC methods were proposed to remove all the silent intervals in speech sample. Experiment are carried out using Linear Predictive Coding (LPC) and Dynamic Time Warping (DTW) for isolated digit recognition in Bahasa Malaysia. The technique without silent removal LPC-DTW gives a recognition accuracy of 98.28%. With detection and removing of silent interval, both technique IE-LPCDTW and LTC-LPC-DTW gives a recognition accuracy of 98%. The system then are applied for training and testing for connected digit recognition. The segmentation of input string of the digits are carried out using IE and LTC techniques. Connected digit recognition using IE-LPC-DTW had 93.3% digit accuracy and 78% digit string. However using LTC-LPC-DTW the performance decreased to 93.2% and 77.7% respectively.