Exploiting forced alignment of time-reversed data for improving HMM-based handwriting segmentation