(A) study on the use of statistical information in speech recognition based on HMM

Abstract

The need for a speaker-independent large vocabulary speech recognition system has been grown due to its large application area. Although the phoneme-level HMM has been widely used as an efficient algorithm for large vocabulary, the performance of the phoneme-level HMM-based recognition system has to be improved more for practical use. For this reason, a new HMM parameter estimation algorithms are proposed to improve the recognition accuracy, and an efficient pre-classification algorithm is proposed to reduce recognition time. In order to estabilish a benchmark performance performance, a phoneme-level HMM-based recognition system is first implemented as a baseline system. And then, the performance of the baseline system is improved. In this work, we focus mainly on the continuous HMMs and the semi-Continuous HMMs to improve the recognition accuracy for isolated Korean words when only insufficient training data are available. First, in order to model temporal changes in spectra, we propose a modified HMM with nonparametric state duration probability and state duration-dependent observation probability to model state transitions and to have accurately temporal structures and timing informations. Our modeling assumption is essentially based on the fact that the temporal changes and the acoustic effects of timing differences in the spectra characterize the time-varying vocal tract, and consequently play an important role in human perception. To model transitions and state durations, and to consider the temporal structures more accurately, we use not only the transition probability, but also a set of state-duration probability combined with state duration-dependent observation probability. Second, the HMM-based speech recognition system uses a training algorithm, which adjusts parameters to obtain an approximation to the maximum-likehoodestimates(MLE) of HMM parameters. The MLE training algorithm does not attempt to maximize the recognition rate on the training data,...