A pattern matching approach to connected-word recognition for Chinese

Abstract

A speaker-dependent pattern-matching approach to connected word recognition for Chinese is presented. First, a method of adaptive energy normalization is applied to the speech spectrum, and a sound stimulus parameter is used to compress the normalized spectrum. Then, using a set of isolated word tokens as the reference patterns, a simplified dynamic programming-based matching strategy using a fast dynamic-time-warping alignment procedure is described. Experimental results for the recognition of a Chinese digit string (of unknown variable length from 2 to 5 digits), for two kinds of pronunciation (in standard Chinese and in the way used in telecommunication), are given. The correct string recognition rate is 96.8% and 97.6%, respectively.<>