A Comparative Study of Several Dynamic Time-Warping Algorithms for Connected-Word Recognition

C. Myers, L. R. Rabiner

Bell System Technical Journal · 1981 · 419 citations · 18 references

Concepts

TL;DR

Multiple algorithms—two-level dynamic programming, sampling, and level-building—have been proposed for time-aligning test patterns with concatenated reference patterns in automatic connected-word recognition. The study examines the theoretical differences and similarities among these algorithms. The authors conduct an experimental comparison of the algorithms on a connected-digit recognition task. The results indicate that the level-building algorithm outperforms the two-level dynamic programming and sampling algorithms in typical applications.

Abstract

Several different algorithms have been proposed for time registering a test pattern and a concatenated (isolated word) sequence of reference patterns for automatic connected-word recognition. These algorithms include the two-level, dynamic programming algorithm, the sampling approach and the level-building approach. In this paper, we discuss the theoretical differences and similarities among the various algorithms. An experimental comparison of these algorithms for a connected-digit recognition task is also given. The comparison shows that for typical applications, the level-building algorithm performs better than either the two-level DP matching or the sampling algorithm.

References

18