Confidence Measure and Rejection Based on Correctness Probability of Recognition Candidates

摘要

We propose a confidence measure expressing the correctness probability of recognition candidates, to be applied to a speech dialog system using a dialog strategy based on the degree of confidence of speech recognition candidates. We assume a function expressing the correctness probability of the feature parameters associated with the recognition candidate, and its parameters are estimated by using the square error of the correct (1) and incorrect (0) data as the evaluation function. For two feature quantities, namely, the likelihood ratio of the recognized word and syllable-concatenated model, and the variance of the syllable duration inside a word, the correctness probability can be expressed by a sigmoid function. The correctness probability can also be expressed for a combination of these two feature parameters. Rejection experiments involving evaluation of the degree of discrimination between the correct and incorrect data show that the degree of discrimination is higher when the two feature parameters are combined than when they are used individually.