DECLINE IN ACCURACY OF AUTOMATIC SPEECH RECOGNITION AS A FUNCTION OF TIME ON TASK - FATIGUE OR VOICE DRIFT

Recognition accuracy of speech recognition devices tends to decline during an extended period of continuous use. Although this deterioration in performance is commonly acknowledged, there has been little systematic observation of the phenomenon, and no clear account of its causes is available. The aim of the present study was to provide some indication of the magnitude and time course of this decline in performance, and to clarify the nature of underlying changes in speech behaviour. There experiments are described. Experiment 1 confirmed that there is a fall-off in recognition accuracy during a half-hour session of a data entry task, and that this occurs for both naive and practised subjects. In Experiment 2, no recovery was observed in recognition performance when short rest breaks were scheduled, indicating that vocal fatigue was not a major factor. The effects of template retraining in mid-session were investigated in Experiment 3. This procedure was found to be effective in restoring recognition accuracy, and the retrained templates were relatively robust. The implications of these findings for operational use of speech recognition devices are briefly discussed. For most applications, one-off template retraining is seen as a more appropriate solution to the problem of voice drift than more complex solutions based on adaptive templates. © 1992.