In a previous paper ((12) Simpson, 1985) it was reported that recognition accuracy for an isolated word, speaker-dependent recognizer, was severely degraded under various levels of pilot task-induced stress. The speech data were collected as part of a larger study of alternative cockpit control and display modes for the LHX mission ((14) Voorhees and Bucher, 1985). In the study reported here, a connected word, speaker-dependent recognizer by the same vendor was used. There was no statistically significant effect on recognition accuracy for task-induced stress, at least for the levels of stress induced in this experiment. However, different types of errors occurred with the connected word recognizer compared to the isolated word recognizer. This paper describes the characteristics of the pilots' speech under stress which were associated with the different types of recognition errors for the two types of recognizers. It is suggested that these human speech performance characteristics must be considered in any implementation of speech recognition in rotorcraft.