论文部分内容阅读
语调短语的自动预测是影响合成语音的可懂度和自然度的关键因素之一。该文提出了一种结合分类与回归树(classification and regression tree,CART)和基于转换的错误驱动的学习(transformation-based error-driven learning,TBL)算法的混合算法,从无限制的文本中预测语调短语边界。该方法根据CART模型的指导,自动生成TBL算法的规则模板,从而减少了规则学习过程中的人工参与。实验结果表明:基于该混合算法的语调短语预测的F-score达70.0%。自动生成的TBL模板不仅能较好地替代手工模板,而且在与手工模板一起使用时,也能为其提供有益的补充。
Auto-prediction of intonation phrases is one of the key factors that affect the intelligibility and naturalness of synthesized speech. This paper proposes a hybrid algorithm that combines classification and regression tree (CART) and transformation-based error-driven learning (TBL) algorithm, which predicts from unrestricted text Intonation phrase boundaries. According to the guidance of the CART model, this method automatically generates the rules template of the TBL algorithm, thereby reducing the manual participation in the rule learning process. The experimental results show that the F-score of intonation prediction based on the hybrid algorithm reaches 70.0%. The automatically generated TBL template not only provides a good alternative to manual templates, but also provides a useful complement when used with manual templates.