Additional use of phoneme duration hypotheses in automatic speech segmentation

In this paper, we describe a new approach for speaker independent automatic phoneme alignment. Typical algorithms for this task use only phoneme-to-frame similarity measures which are somehow maximised or minimised. In addition to such similarity measures, we use phoneme duration hypotheses generated by the speech synthesis system HADIFIX [1]. For algorithms based on dynamic programming, it is difficult to use these duration hypotheses, so we create a cost-function consisting of phoneme-to-frame and segment-to-duration hypotheses similarity measures and minimise this cost-function by a Genetic Algorithm. The results show that the accuracy of automatically determined phoneme boundaries increases. This accounts especially for speakers not used in the training phase.