Weighting Phone Confidence Measures for Automatic Speech Recognition

One of the most useful applications of Confidence Measures (CMs) in Automatic Speech Recognition systems is early detection of incorrect recognition hypotheses. A purely acoustic basis for such a CM is particularly important when tracking errors resulting from Out of Vocabulary speech, background noise or keyword substitution. A commonly taken approach is to compute scores on subword units of the hypothesized words and combine them in a word score. This paper investigates the assumption that some subword types contain stronger distinctive properties than others. Therefore, their scores ought to have a higher contribution in the eventual word scores. Experiments in a connected digit recognition task showed a relative Confidence Error Rate improvement of 6% on word level and 11% on sentence level in comparison to the baseline CM, with equal contribution of the phone confidence scores.