Investigation of Confirmation Strategies for Speech Recognition Applications

Abstract

Guidelines for speech user interfaces generally promote the use of delayed confirmation in speech recognition applications that require users to provide multiple elements of information. In our initial investigation of a simple delayed confirmation method, we discovered a significant design flaw (requiring users to review a fairly large amount of correct input multiple times when asked to correct two errors). To avoid this flaw, we designed two new methods. In one (Serial Collection/Correction), users named an item that needed correction, then made that correction before naming the next item to change. In the other (Batch Collection/Correction), users named all items that needed correction first, then changed the named items in sequence. An experiment comparing the methods indicated no significant difference in user preference, but significantly fewer memory errors occurred when using the Serial Collection/Correction method. Thus, we recommend the use of the Serial Collection/Correction method. We also recommend including an option to replay the information if a user has remembered the need to change an additional item, but does not remember which item to change.