Based on a speaker-independent continuous speech recognizer working with a conventional bi-gram language model our approach integrates a finite state grammar into the recognition process. For syntactically structured tasks this grammar is represented as bigram language model. Thus the implementation of the recognition process remains unchanged. We report the results of an evaluation on a German database consisting of 1500 strongly structured commands. Test set perplexity is nearly halved and a significant higher recognition rate is reached without a notable delay in the recognition process.