Estimation of word confidence using unscented transform for the rejection of misrecognized words

Confidence measure (CM) is generally used for the rejection of misrecognized words in automatic speech recognition (ASR). Common ASR systems test recognized words with anti-models, especially anti-phone models. That is, they calculate a word confidence with a nonlinear function, which weight nonlinearly phone-level CMs. Word confidence estimation is important in word-independent isolated word recognition, for the fixed threshold can cut out some words excessively. The statistics of word-level confidences are not consistent. In this paper we present a simple and efficient method for estimating word confidence using the unscented transform with the statistics of phone-level confidence values. Our experimental results show that the unscented transform with tri-phone-based CM statistics can estimate word confidence reasonably.