Utterance Verification Based on the Likelihood Distance to Alternative Paths

Utterance verification tries to reject incorrectly recognised utterances. For this purpose the probability of an error is often estimated by single confidence measure (CM). However, errors can have several different origins, and we argue that notion must be reflected in the design of the utterance verifier. In order to detect both in-vocabulary substitutions and out-of-vocabulary word errors, we compute CMs based on the log-likelihood distance to (1) the second best recognition result and to (2) the most likely free phone string. Experiments in which different CMs were combined in different ways in the recognition of Dutch city names show that Confidence Error Rates [8] are reduced by 10% by combining the CMs using a classification and regression tree instead of a linear combination with a decision threshold.