面向语音合成的藏语音素切分算法研究

Study on Tibetan Phoneme Segmentation Algorithms Facing Speech Synthesis

文章通过采用两种方法对藏语语音合成语料库中的语音进行音素切分:一种是基于单音素HMM模型的自动切分方法,一种是传统的人工切分方法,并通过实验分析了自动切分与人工切分方法的准确率程度.实验结果表明:在构建语料库时,前者有助于缩短建库周期,尤其对于大语料库的建立会有明显的优势.这种方法既节省了切分与标注的大量时间和人力成本,又提高了语音语料库标注信息的精确度和一致性.

This paper adopted two methods being used for phoneme segmentation for Tibetan speech synthesis corpus：one was based on single phoneme HMM model automatic segmentation;the other was the traditional manual segmentation way.The accuracy degree between automatic and manual segmentation was analyzed through the experiments.The results of experiment showed that the automatic segmentation is helpful for shortening the cycle duration in building corpus process,especially for the establishment of large corpus.A lot of time for segmentation and labeling was reduced,the accuracy and consistency of speech corpus labeling information has been improved.