Speech variability effects on recognition accuracy associated with concurrent task performance by pilots

Abstract

Pairs of pilots performed an aircraft warning classification task using an isolated-word speaker-dependent speech recognition system to record their responses. Task-induced stress was manipulated by means of different scoring procedures for the warning classification task and by inclusion of a competitive manual control task (Voorhees and Bucher 1984, Voorhees and Bucher 1985). High quality tape recordings were made of pilots’ speech spoken to the recognizer while performing under the different types of task loading. This speech was then analysed spectrographically for timing differences in the pilot's manner of speaking as a function of task loading, and of recognition accuracy associated with different types of task loading. Recognition errors were recorded by type for an isolated-word speaker-dependent system and by an off-line technique for a connected-word speaker-dependent system. While recognition errors increased with task loading for the isolated word system, there was no such effect for task loading when tape recordings of the same utterances were recognized by a connected word system. The results suggest areas for further improvements to speaker-dependent recognition algorithms and provide predictions of the types of recognition errors that can be expected with current recognition technology used by pilots under moderate to high levels of task loading.