As part of its internal research program, MITRE has developed an experimental speech recognition system for small-vocabulary simple-grammar applications. This system, called Hark, has three parts: a signal processor for extracting feature vectors from the acoustic signal, a neural network for classifying the feature vectors, and a dynamic programming module for decoding the utterance. Hark is speaker independent and handles continuous speech with pauses after sentences. This paper provides an overview of the Hark system.