A slower phrase in spontaneous conversational speech is caused by emphasis, thinking during speaking and so on. To include such useful information with man-machine communication, we investigate a method to detect local slower phrase from time sequence of mora duration in Japanese dialog speech. At first we prepare speech samples, which contains phrases slowed considerably. Then the flow of the process to obtain phrase averaged mora duration is explained. In this method, speech period, mora boundaries and phrase boundaries are obtained from acoustical features. A threshold is applied to phrase averaged mora duration. An experiment to detect a local slower phrase is carried out. The slowed phrases are detected with high recall rate.