On the Use of Dynamic Time Warping for Word Spotting and Connected Word Recognition

* The work presented here is based, in part, on the MS thesis, "A Comparative Study of Several Dynamic Time Warping Algorithms for Speech Recognition," by C. S. Myers, MIT, April 1980. 303 Itakura, and White and Neely has shown that DTW algorithms are an effective method of time registering patterns in isolated word recognition systems. Bridle and Christiansen and Rushforth have studied the applicability of DTW algorithms to word spotting, and recently, Sakoe, Rabiner and Schmidt, and Myers and Rabiner, have successfully applied dynamic time-warping techniques to connected digit recognition. A great deal of work has been done in the area of performance evaluation of the various DTW algorithms as applied to discrete word recognition. " However, the effects of the DTW parameters on the overall performance of the algorithm for either word spotting or connected word recognition are not as well understood. The purpose of this paper is to discuss several proposed methods of applying DTW algorithms to word spotting and connected word recognition, and to study some of the factors which determine the performance of these algorithms. The organization of this paper is as follows. In Section II we review the basic dynamic programming method of time alignment and show how it may be used efficiently in either a word spotting or a connected word recognition problem. We describe, in detail, two different DTW algorithms for which we have performed extensive evaluations. Section III contains a description of the experiments which we performed to evaluate the performance of the different DTW algorithms and the effects of the parameters associated with them.