An Efficient Word Confidence Measure Using Likelihood Ratio Scores

Abstract

This paper describes an efficient method to perform word confidence measures in an automatic speech recognition system. The confidence measure is computed during the decoding phase and is based on likelihood ratios between the top hypotheses that reach a word node. Experiments were carried out on a digit database with a connected-digit recognizer. The results show that this method outperforms word-graph confidence measure with a special grammar and is worse with a word loop grammar. Because the proposed confidence measure is evaluated with only one pass, it is very efficient and can be applied with advantage in small or medium vocabulary recognizers, with low computational resources.

Resources

Details