We have developed a Bayesian classifier to determine whether syllables in connected Thai speech are weakly or strongly stressed by using five acoustic parameters: syllable rhyme duration, mean F0, F0 standard deviation, mean energy, and the standard deviation of the energy. With speaker-dependent data normalization, we achieved a classification accuracy of 99%. The classification accuracy drops to 96% when we used speaker-independent normalization. We have also developed prosodic constraints that can use this stress information to syntactically disambiguate a class of ambiguous sentences that arise from the use of compounding in spoken Thai.