多项式回归的汉语时长预测模型

时长信息是韵律的重要组成部分,对于语音合成的自然度和可懂度都有不可忽视的作用。时长预测是建立对时长有影响的韵律环境与自然语流中音段时长的对应关系。本文引入了统计学中etasquared 的概念研究汉语中韵律环境因素对时长的影响,设计了残差算法定量分析属性之间的交互作用,由此建立了多项式回归的汉语时长预测模型。实验结果表明,使用5～6 个韵律属性基本上就能够建立比较相关的对应关系,和使用同样韵律属性的Wagon 回归树的效果相比有明显的优势。

Polynomial Regression Model for Duration Prediction in Mandarin

Duration information is an essential part of speech prosody , and plays a critical role in improving the naturalness and understandability of synthesized speech. Duration modeling is to establish a mapping relationship between the prosodic environment and the final duration engendered in natural speech. In this paper , we first study the effect of prosodic features on segmental duration by introducing a statistical concept —eta squared , then choose more forceful prosodic features and design an algorithm to quantify the interaction among them , and finally bring forward the method of determining the duration model using a polynomial equation and obtain the coefficients through non - linear regression. Our research work indicates that 5 or 6 prosodic features might by and large assist a close and accurate mapping between prosodic environment and perceived duration. Compared to Wagon tree method , this method has undeniable merits.