Abstract: Glottal stops in cleft palate speech of Putonghua are abnormal productions in which closures in the glottis substitute typical obstructions along the vocal tract(e.g.,alveolus,lips).In this study,we investigated the spectral characteristics of glottal stops that patients intended to substitute for initial unaspirated alveolar stops.The first four moments were calculated for the burst and the subsequent early portion of the stop consonant.The results show that the first moment(i.e.,centroid) was lower for glottal stops than for normal alveolar stops.That is,there was less acoustic energy in the high frequencies for glottal stops than for alveolar stops,due to the larger cavity anterior to the closure at the glottis than at the alveolus.We also found that the spectral energy of glottal stops spread over a wider frequency range,indicated by the greater second moment (i.e.,variance or standard deviation).Positive skewness showed that the spectral distributions were asymmetric with the head towards low frequencies and the tail towards high frequencies.As these spectral moments were taken as variable candidates,a logistic regression model was adopted as the binary classifier to determine whether an unaspirated alveolar stop consonant was a glottal stop.Acceptable classification accuracy was obtained after k-fold cross-validations and an optimal choice of a few spectral moments.This study suggests that it is feasible to implement a system prototype to automatically detect glottal stops in cleft palate speech by integrating the classifier with the measurement of burst time.