Predicting sound segment duration in connected speech: an acoustical study of brazilian portuguese