Word spotting based ona posterior measure of keyword confidence

Abstract

In this paper, an approach of keyword confidence estimation is developed that well combines acoustic layer scores and syllable-based statistical language model (LM) scores. An a posteriori (AP) confidence measure and its forward-backward calculating algorithm are deduced. A zero false alarm (ZFA) assumption is proposed for evaluating relative confidence measures by word spotting task. In a word spotting experiment with a vocabulary of 240 keywords, the keyword accuracy under the AP measure is above 94%, which well approaches its theoretical upper limit. In addition, a syllable lattice Hidden Markov Model (SLHMM) is formulated and a unified view of confidence estimation, word spotting, optimal path search, and N-best syllable re-scoring is presented. The proposed AP measure can be easily applied to various speech recognition systems as well.