Detail publikace

Abstrakt

This paper wishes to contribute to the research activities in the field of detection of incorrectly recognized words during automatic speech recognition. In this paper we describe our original detection technique of incorrectly recognized words and compare it to the one we previously used. Our speech recognition system is based on a speaker-independent continuous density Hidden Markov Model and a dedicated phone sequence recognizing network, which we call a mumble model. The incorrectly recognized word detection method is fully statistically based and uses a MLE (EM) approach to estimate the parameters of statistical models determining the confidence of automatic speech recognition results. We use two kinds of models: the acceptation model and the rejection model. According to their names, the acceptation model corresponds to the words that were correctly recognized, while the rejection one represents all other words, i.e. the misrecognized and out of vocabulary words, they cannot be recognized at all.