Dynamic programming method for fine-tuning the boundary points in automatic segmentation of speech

The important element of today's speech systems is the set of recorded wavefiles annotated by a sequence of phonemes and boundary time-points. As the manual segmentation of speech is a very laborious task, there is the need for automatic segmentation algorithms. However, it was observed that common HMM-based methods are prone to systematical errors. Thus, some boundary refinement approaches were introduced. In this paper we combine two sources of information: boundary error distribution and an acoustic observation distribution, in a single dynamic programming approach.