Likelihood normalization using an ergodic HMM for continuous speech recognition

In recent speech recognition technology, the score of a hypothesis is often defined on the basis of HMM likelihood. As is well known, however, direct use of the likelihood as a scoring function causes difficult problems especially when the length of a speech segment varies depending on the hypothesis as in word-spotting, and some kind of normalization is indispensable. In this paper, a new method of likelihood normalization using an ergodic HMM is presented, and its performance is compared with those of conventional ones. The comparison is made fr om three points of view: recognition rate, word-end detection power, and the mean hypothesis length. It is concluded that the proposed method. gives the best overall performance.