Confidence Measures

Abstract

The majority of ASR systems achieve recognition rates that are well below those achieved by humans. Especially with increasing vocabulary size, the probability that words are confusable increases, thus making it more difficult to recognise the correct word. Furthermore, many commercially important speech recognition tasks require the ability to understand spontaneous rather than isolated speech, which is an even bigger problem. Although this provides a user friendly user interface, it poses a number of additional problems, such as the handling of out of vocabulary (OOV) words, disfluencies and acoustical mismatch. And unaware of the technology limitations, users expect the system to work properly, even if their utterance includes hesitations, false starts and sounds like uhm‘s and ah‘s.