Refining Segmental Boundaries using Support Vector Machine

High accuracy phonetic segmentation is critical to achieve good quality in concatenative speech synthesis. However, the processing and inspection of a large amount of recorded speech will become a labor-intensive and error-prone job. In this paper, a post-refining method based on support vector machines (SVMs), is proposed for auto-segmentation of speech data. Our baseline system is based on a set of hidden Markov models (HMMs). This system performs forced alignment of speech data and phonemic transcription of corresponding text. A de-biasing algorithm first refines initial boundary estimates. SVM models are then used for more refinement of de-biased boundaries. Subsequently, a LBG vector quantization algorithm is used to reduce the amount of speech for training SVM models. This leads to a considerable decrease in necessary time to train SVM models. We achieved a performance of 94.3% for segmentation of phoneme boundaries with less than 15 ms deviation from hand labeled boundaries