This paper describes an approach in which phrase accent information is used for dialogue act recognition in German spontaneous speech. This application is an example of how automatically computed prosodic information can be used in automatic speech recognition. Usually the important intention conveyed by an utterance is found in the focused area, which is often accentuated. When all the words of an utterance are used for dialogue act classification, the best result is achieved only if all probabilities (e.g. of n-grams) are known. In real life applications this not the case. Because utterances can be very similar to one another, but belong to different dialogue act classes, it may be possible to distinguish the classes on the basis of characteristic words. For this reason dialogue act classification is often based on keyword detection. The selection of keywords is crucial. Better recognition relies on better chosen keywords. This paper shows how keyword selection can be improved by using two additional information sources: lexical POS information and prosody. POS and prosodic information is used to build subsets of the vocabulary to improve recognition. Experiments are conducted on a sub-sample of the VERBMOBIL corpus. The aim is to distinguish between four dialogue act sub-classes of the general class SUGGEST.