Abstract

This thesis is concerned with the rhythmical structure of speech under the influence of alcohol. All analyses presented are based on the Alcohol Language Corpus, which is a collection of speech uttered by 77 female and 85 male sober and intoxicated speakers. Experimental research was carried out to find robust, automatically extractable features of the speech signal that indicate speaker intoxication. These features included rhythm measures, which reflect the durational variability of vocalic and consonantal elements and are normally used to classify languages into different rhythm classes. The durational variability was found to be greater in the speech of intoxicated individuals than in the speech of sober individuals, which suggests, that speech of intoxicated speakers is more irregular than speech of sober speakers. Another set of features describes the dynamics of the short-time energy function of speech. Therefore different measures are derived from a sequence of energy minima and maxima. The results also reveal a greater irregularity in the speech of intoxicated individuals. A separate investigation about speaking rate included two different measures. One is based on the phonetic segmentation and is an estimate of the number of syllables per second. The other is the mean duration of the time intervals between successive maxima of the short-time energy function of speech. Both measures denote a decreased speaking rate in the speech of intoxicated speakers compared to speech uttered in sober condition. The results of a perception experiment show that a decrease in speaking rate also is an indicator for intoxication in the perception of speech. The last experiment investigates rhythmical features based on the fundamental frequency and energy contours of speech signals. Contours are compared directly with different distance measures (root mean square error, statistical correlation and the Euclidean distance in the spectral space of the contours). They are also compared by parameterization of the contours using Discrete Cosine Transform and the first and second moments of the lower DCT spectrum. A Principal Components Analysis on the contour data was also carried out to find fundamental contour forms regarding the speech of intoxicated and sober individuals. Concerning the distance measures, contours of speech signals uttered by intoxicated speakers differ significantly from contours of speech signals uttered in sober condition. Parameterization of the contours showed that fundamental frequency contours of speech signals uttered by intoxicated speakers consist of faster movements and energy contours of speech signals uttered by intoxicated speakers of slower movements than the respective contours of speech signals uttered in sober condition. Principal Components Analysis did not find any interpretable fundamental contour forms that could help distinguishing contours of speech signals of intoxicated speakers from those of speech uttered in sober condition. All analyses prove that the effects of alcoholic intoxication on different features of speech cannot be generalized but are to a great extent speaker-dependent.