Errors? Not too worrisome. Exploring the effects of errors in speech transcription

This thesis presents two user studies that investigate how the errors that can occur during speech recognition affect users' text entry performance and their experience. For this work, I used a speech recognition system that injects believable errors in a controlled manner, and where users could fix errors simply by re-speaking a small part of their original utterance. Participants were asked to transcribe a set of phrases using the system, either with or without the insertion of errors, In the first study, I injected up to 33% errors, but saw no substantial differences. Yet, participants consistently commented on the used phrase set, which did not correspond well with spoken English. Thus, I created a novel phrase set based on spoken phrases. In the second study, I used this phrase set and inserted errors into the speech recognition results with either 25% or 50% probability. The results showed that inserting errors in the speech recognition system had a significant effect on participants' perceived mental workload. In addition, I found that inserting errors increased the number of errors users made during the task. More importantly, I discovered that the more erroneous the system was, participants were less likely to detect errors. The new phrase set for spoken English also performed substantially better than the one used in the first set. Finally, according to my findings, users have a fairly high tolerance towards errors encountered in speech transcription.