In this paper we analyze postposition drops in nominal phrases of spoken Japanese. This analysis is aimed at being applied to syntactic constraints within a speech recognition module of a speech translation system. Speech recognition often fails to recognize short words. Among syntactic categories, postpositions belong to such a group. In machine translation of Japanese, postpositions are often used as a processing trigger. However, if a postposition is unspoken, this approach fails. To overcome this problem as well as reflect actual usage, we have studied postposition drops in our dialogue database. We first retrieved all nominal phrases without any postposition. We then studied the nominal phrases in which the dropped postpositions can be naturally recovered and those for which postposition recovery is very difficult. Further, we investigated ways of predicting the recoverable postpositions that require little calculation. Finally, we discuss the possibility of applying the analysis to syntactic constraints within speech recognition.