Using and evaluating new confidence measures in word-based isolated word recognizers

In this paper a method for detecting out of vocabulary words in isolated word recognizers is introduced, our method utilized new kinds of confidence measure. After recognition task was completed and consequently confidence measure was extracted, a classifier would accept or reject result of recognition task using this CM.We used two different kinds of confidence measure where for extracting each one a different information source was used. Amount of competition between hypotheses through the recognition task was used for extracting first CM. The second one was extracted using information about manner of distribution of feature vectors in the states of winner HMM model. Both of these CMs were used separately and their ability for detecting out of vocabulary words was measured. First CM results a hit rate of 87% and false alarm of 12% and the second one results 86% and 19% consequently.