Dynamic Time Warping with Boundaries Constraint Relaxation

Conventional Dynamic Time Warping (DTW) assumes exact knowledge of the boundaries of both reference and test sequences. However, the output of practical end point detectors is inaccurate, especially with noisy input. This results in a severe deterioration of the accuracy of isolated word recognition. This work presents a modified DTW algorithm in which the boundaries constraint is relaxed so as to reflect endpoint detection errors. The effect of this modification on the dynamic programming optimization is analyzed and experimental results axe presented to demonstrate the capability of the algorithm to overcome endpoint errors.