Pseudo-continuous hidden Markov modeling for automatic speech recognition

The methods explored try to eliminate apparent framing artifacts by allowing variable rates in sampling the speech parameters. Traditional hidden Markov model speech recognition is generally based on a set of parameters which are extracted at discrete intervals. Such an analysis necessitates use of a discrete-transition hidden Markov model in which the underlying states can change only at intervals related to the frame rate of the analysis. The exact locations of the analysis windows can influence the front-end outputs. As a result, inconsistent performance can often be observed in discriminating words which differ only in short duration cues. Methods are explored which circumvent this framing effect by allowing state transitions to occur at each sample. Efficient methods for implementing this strategy are derived, and experimental results are presented, showing the superior performance of the pseudo-continuous transitions hidden Markov models compared to that of conventional discrete transition hidden Markov models