Reduction of minimum word-boundary gap lengths in isolated word recognition

Abstract

In this paper, we describe an isolated-word recognition system which challenges the assumption that only connected word recognition can provide high-speed data entry. In previously available isolated-word recognizers, the entry rate was not only limited by the user's ability to speak words in isolation, but by the requirement to leave relatively large gaps between words in order for the system to use gap length to discriminate word boundary gaps from intra-word stop-gaps. The new isolated-word recognition algorithm greatly reduces the minimum acceptable word-boundary gap length by using correlation scores and dynamic programming string matching to discriminate between the two kinds of gaps. Reliable word recognition results are obtained even with word-boundary gaps which are a fraction of the length of typical stop-gaps, and the effective digit entry rate is about twice that of previous isolated-word recognition systems.