Automatic speech recognition by computer must address two issues to perform reliably in actual recognition environments. The first issue is real-time system performance. The second issue is the effect of background noise and/or speaker stress on recognition performance. A considerable effort has been made to develop speech recognition in tranquil environments. However, speech recognition algorithms formulated in tranquil environments generally perform poorly in adverse environments (background noise and/or speaker stress). In this paper, we propose a real-time recognition system called ICARUS which addresses the effect of background noise/speaker stress. The motivation for our stress compensation scheme is discussed with respect to the variation of speech characteristics spoken in noisy environments. Results of the proposed system are given for several speakers. ICARUS represents a first attempt at real-time speech recognition in adverse conditions. Early performance results, though not consistent over all speakers, show improvement in recognition of as much as +7.1 % for noise free Lombard speech, +15.7 % and +7.1 % for Lombard speech corrupted by additive white Gaussian and non-stationary cooling fan noise respectively.