Sources of Human Error in Data Entry Tasks Using Speech Input

Abstract

Where speech recognition devices are used for entering data, some form of feedback is usually provided so that recognition errors can be detected and corrected. The present experiments were designed to investigate how the accuracy of data entry is influenced by the type of feedback used and by the structure of the error correction dialogue. Experiment 1 compared the effects of visual and spoken feedback delivered either after each item or at the end of the data string. With visual feedback, most failures to achieve correct data entry occurred when subjects failed to detect recognition errors. Spoken feedback was much more closely monitored, but with word-by-word feedback, subjects had difficulty in recovering from mis-recognitions in the error correction dialogue. Experiment 2 was designed to examine possible interactions between presentation and feedback modalities when data strings were entered from memory. Although there was some evidence of an interaction, the effects were small in relation to the much greater disruption of memory caused by correction of recognition errors in midentry. The implications of these findings for the design of feedback and error correction dialogues are discussed.