Segmentation of Malay Syllables in Connected Digit Speech Using Statistical Approach

Abstract

This study present segmentation of syllables in Malay connected digit speech. Segmentation was done in time domain signal using statistical approaches namely the Brandt’s Generalized Likelihood Ratio (GLR) algorithm and Divergence algorithm. These approaches basically detect abrupt changes of energy signal in order to determine the segmentation points. Patterns used in this experiment are connected digits of 11 speakers spoken in read mode in lab environment and spontaneous mode in classroom environment. The aim of this experiment is to get close match between reference points and automatic segmentation points. Experiments were conducted to see the effect of number of the auto regressive model order p and sliding window length L in Brandt’s algorithm and Divergence algorithm in giving better match of the segmentation points. This paper reports the finding of segmentation experiment using four criterions ie. the insertion, omissions, accuracy and segmentation match between the algorithms. The result shows that divergence algorithm performed only slightly better and has opposite effect of the testing parameter p and L compared to Brandt’s GLR. Read mode in comparison to spontaneous mode has better match and less omission but less accuracy and more insertion.