Mobile service robots in human environments need to have versatile abilities to perceive and to interact with their environment. Spoken language is a natural way to interact with a robot, in general, and to instruct it, in particular. However, most existing speech recognition systems often suffer from high environmental noise present in the target domain and they require in-depth knowledge of the underlying theory in case of necessary adaptation to reach the desired accuracy. We propose and evaluate an architecture for a robust speaker independent speech recognition system using off-the-shelf technology and simple additional methods. We first use close speech detection to segment closed utterances which alleviates the recognition process. By further utilizing a combination of an FSG based and an N-gram based speech decoder we reduce false positive recognitions while achieving high accuracy.