Improvement of time alignment of the speech signals to be used in voice conversion

One of the main applications of time alignment is parallel corpus based voice conversion. In the literature, various methods such as dynamic time warping (DTW) and hidden Markov model have been suggested for time alignment of two speech signals. In this paper, we introduce some modifications to DTW in order to decrease the time alignment error. These modifications are refinement, which is done by exerting a threshold, normalization, and comparisons between the preceding and the following frames to make sound correspondence between two different parallel corpus-based speakers’ speeches. Evaluation of this approach which has been done on some corpus sentences indicates a significant improvement of time alignment. At least about 4% and in some cases 15% decrease of error in comparison with DTW has been achieved.