A Study of the Use and Evaluation of Con dence Measures in Automatic Speech Recognition

Con dence measures have been found to be useful for a number tasks within the eld of Automatic Speech Recognition (ASR). For example, the use of con dence measures has been reported in the utterance veri cation, keyword spotting and Out-of-Vocabulary (OOV) word spotting literature. In this report, it is shown that so called 'hybrid Arti cial Neural Network/Hidden Markov Model' (HMM/ANN) systems are well suited to the task of generating con dence measures, due to their ability to provide local phone class posterior probability estimates which may be used to generate con dence measures in a computationally e cient manner. A number of evaluation metrics are also described and the performance of ve con dence measures derived from the ABBOT hybrid HMM/ANN system for the tasks of utterance veri cation and OOV word spotting are evaluated using these metrics. Besides the tasks described above, con dence measures may also be used for tasks such as ltering the acoustics for a number of conditions prior to decoding, guiding the search over the space of alternative decodings and also as diagnostic tools. The use of con dence measures as diagnostic tools can indicate how components of an ASR system can be improved. For example, con dence measures can be used to identify poor pronunciation models and also to guide the search for improved models, where the generation of accurate pronunciation models remains an important challenge in the creation of practical recognition systems.