Note on some factors affecting performance of dynamic time warping algorithms for isolated word recognition

To implement a dynamic time warping (DTW) algorithm for isolated word recognition, a number of factors must be specified. These factors include endpoint constraints, local path constraints, global path constraints, axis orientation, and local distance measure. Although a number of studies have been made to decide how best to choose these factors, several unresolved issues remain. The problems with choosing these factors become more complicated if the word reference patterns are not created from whole words, but instead are built up from subword units, e.g., syllables or demisyllables. In this paper, we consider an isolated word recognizer in which the reference patterns are obtained from a set of demisyllables as specified in a user-supplied lexicon. Different pronunciations of a word are represented by multiple entries in the lexicon. Because of the inherent boundaries in the reference patterns (at each demisyllable junction), it was felt that local constraints could influence performance more than for the standard whole-word case. Further, it was felt that some flexibility in the endpoint path constraint at the end of the word would be helpful, in general, for isolated word recognition caused by high variability at the ending of words with stops. A DTW algorithm was programmed and tested on an 1109-word vocabulary. Results showed small accuracy improvements when the path endpoint was allowed to vary across four frames (of either test or reference pattern). Loosening the local path constraints, however, had a significant degrading effect on the performance.