Automatic Speech Segmentation Based on HMM

This contribution deals with the problem of automatic phoneme segmentation using HMMs. Automatization of speech segmentation task is important for applications, where large amount of data is needed to process, so manual segmentation is out of the question. In this paper we focus on automatic segmentation of recordings, which will be used for triphone synthesis unit database creation. For speech synthesis, the speech unit quality is a crucial aspect, so the maximal accuracy in segmentation is needed here. In this work, different kinds of HMMs with various parameters have been trained and their usefulness for automatic segmentation is discussed. At the end of this work, some segmentation accuracy tests of all models are presented. data cannot be segmented manually any more, so it is necessary to use some kind of automatic segmentation in this case. Another example of automatic segmentation necessity can be a data preparation for the initialization phase of a HMM training.