Word Spotting Based on a posterior Measure of Keyword Confidence

In this paper, an approach of keyword confidence estimation is developedthat well combines acoustic layer scores and syllable-based statisticallanguage model (LM) scores. An a posteriori(AP) confidencemeasure and its forward-backward calculating algorithm are deduced. Azero false alarm (ZFA) assumption is proposed for evaluating relativeconfidence measures by word spotting task. In a word spottingexperiment with a vocabulary of 240 keywords, the keyword accuracyunder the AP measure is above 94%, which well approaches itstheoretical upper limit. In addition, a syllable latticeHidden Markov Model (SLHMM) is formulated and a unified view of confidenceestimation, word spotting, optimal path search, and N-best syllablere-scoring is presented. The proposed AP measure can be easily appliedto various speech recognition systems as well.