基于决策树CART选择拼接单元的英语语音合成

English Speech Synthesis Using CART-based Unit Selection

以英语文语转换系统的开发为背景,采用基于大语料库的拼接语音合成方法进行英语语音合成。就英语多音节和无限词汇的特点,选用了3种不同长度的拼接单元:单词,音节,phone。引入了决策树CART(classification and regression tree)方法对大语料库中的语音单元进行预选,并设计了相应的单元选择算法。实验表明,利用该方法能得到清晰自然的合成效果,并且提高了单元选择的效率。

This paper carries out English speech synthesis by large-corpus based concatenation method in the English TI＇S system. Three kinds of concatenating unit of different length are selected which are word, syllable and phone, taking into account the characteristic of multi-syllable and infinite-vocabulary in English. A CART-based approach is introduced to unit pre-selection and the corresponding unit selection algorithm is designed. Test indicates that this method can acquire clear and natural synthesis speech and can improve the efficiency of unit selection.