| 摘要 |
<p>A continuous speech recognition system is described in which false recognition due to an unnatural matching path is reduced. In the system a similarity measure and a path information are developed at each time point (i, j) in the order of monotonous increase of i and j based on the similarity measure and path information obtained up to the time point (i, j) by making use of Dynamic Programming, where i and j designate time points at the time axes of an input pattern and a continuous reference pattern, respectively. The input pattern is given in a time series of feature vectors for an input speech and the continuous reference pattern is given in a time series of feature vectors for connected combinations of several reference word patterns. The similarity measure is given by an accumulated distance between the feature vectors of the input pattern and the reference word patterns. The path information represents a start time point of a word unit of the continuous reference pattern. Circuitry is provided for setting the minimum similarity measure obtained for a word unit as an initial value for a next subsequent word unit. Circuitry is provided for checking whether or not the path information at the time point (i, j) is within a predetermined window restriction. Circuitry is provided for excluding a time point outside the window restriction from development of the similarity measure and path information. Finally, circuitry is provided for deciding recognition results of the input speech on the basis of the matching path showing the minimum similarity measure at the end time point of the input pattern.</p> |
