Confidence measurement techniques in automatic speech recognition and dialog management

Reliable confidence measures are essential to the basis of decisionmaking for enriching human-machine speech interaction with necessary intelligence in ergonomic dialog management. In addition to a survey of the state of the art in confidence measurement, this work also provides classification of methods derivated from several points of view and describes possible fields of application. The thesis includes comparative evaluation results of different computation algorithms which apply posterior probability as the hypothesis confidence measure in HMM-based speech recognition. The key contribution of the dissertation is the description of several utilization techniques that rely on confidence measurement and are intended to enhance the performance of speech recognition systems. A new confidence-guided approach is presented to control the pruning of the Viterbi search process dynamically by taking variable search quality into consideration to fit time-variant requirements. The thesis explores dialog management strategies and several aspects of improving user acceptance in speech-based applications by the use of confidence measurement.