Effect of Operator's Task Load on Performance of a Voice Recognition System

Abstract

G. K. POOCK AND J. W. ARMSTRONG Naval Postgraduate School Canadian Forces Harris, North, and Owens ( 1 ) observed a statistically significant ( p < .05) increase in error rate on a voice-recognition device with increased task load on five operators. Skriver's research ( 3 ) suggested a similar relationship. The present study is the first of a sequence investigating this phenomenon in detail. A discrete utterance voice recognition system (Model T600) from Threshold Technology, Inc. was used with a 50-word vocabulary. The vocabulary included 20 words commonly used in Command and Control environments and 30 rhyming words. All vocabulary words were one or two syllables long. This particular vocabulary was chosen to increase rates of recognition error and to make changes in these rates easily detectable. Subjects individually trained on the TGOO and then TGOO recognition of the vocabulary was assessed under each of 4 conditions of operator's task load; each condition lasted 5 min. In the Control condition with no loading, subjects simply repeated vocabulary words they heard via a headset. In each of the other three conditions subjects similarly repeated vocabulary words but simultaneously performed a psychomotor task on a General Dynamics Response Analysis Tester (RATER). The three RATER tasks were Delay 0, Delay 1, and Delay 2; a presentation rate of one symbol every 1.5 sec. was used. Subjects were instructed that word repetition for recognition by the T6OO had priority. Each of the 24 subjects was nested within an experience level factor as either non-experienced or semi-experienced with voice input. Incorrect recognitions divided by total trials provided an error rate which was examined for the first and second 2.5 min. of each 5-min. trial. Thus, a three-factor nested-factorial design was employed. A parametric analysis of variance was employed since the data met the required assumptions. The effect of the operator's cask loading on recognition error rate was significant (F8.w = 4.35, p < .01) with means, respectively, being: Control = 10.77% (SD = .62), Delay 0 = 13.18% (SD = .74), Delay 1 = 13.14% (SD = .77), and Delay 2 = 13.60% (SD = .70). A range test showed that the error rate was 20% less ( p < .01) under the Control condition than under the three psychomotor loading conditions which were all the same. Recognition error rate was greater in the second half of the 5-rn~n trials (K,E = 14.69, p < .001) wlth the means, respectively: first 2.5 min. = 11.73% (SD = .97), second 2.5 min. = 13.61% (SD = 1.0). Experience level was not sign~f~cant or were the interactions. The average rate of recognition error on the 20 Command and Control words was 2.91% and equivalent to the Command and Control results of Poock ( 2 ) . The average rate of recognition error on the rhyme words was 19.18%. The lack of effect of s u b jects' experience may result from the most experienced subject's having had only 12 hr. of prior work.