【作者】 宋阳； 凌震华； 戴礼荣；

【Author】 Yang Song,Zhen-Hua Ling,Li-Rong Dai National Engineering Laboratory for Speech and Language Information Processing,University of Science and Technology of China,Hefei 270027,China

【机构】 中国科学技术大学语音及语言信息处理国家工程实验室；

【摘要】 近年来提出的基于隐马尔科夫模型的单元挑选语音合成方法,较好的解决了传统拼接合成中存在的依赖较多人工干预以及合成效果不稳定性的问题。但在该方法中,综合不同声学统计模型度量时使用的模型权值无法自动训练获得,且人工优化较为困难。本文提出了一种基于合成质量预测的模型权值优化方法。该方法首先收集较少的人工测听结果并采用多元自适应回归样条构建针对不同权值下合成语音质量的预测模型,然后基于该预测模型利用模式搜索算法自动搜索最优权值。实验证明该方法可以有效优化模型权值并改善合成语音的自然度。更多还原

【Abstract】 HMM based unit selection speech synthesis method has been proposed in recent year.This method can improve the automation of system construction and stability of synthetic speech quality.However,the weights combining with different acoustic statistic models can’t be obtained through the automatic training and it is difficult to tune them manually.Therefore,an approach based on synthesis quality prediction is proposed in the paper for optimizing these model weights.First,subjective evaluation results are collected and a prediction model based on MARS is constructed to predict the quality of synthetic speech using different model weights.Second,a pattern search algorithm is applied to search for the optimal weight.Our experiments indicate that this method can optimize the model weights and improve the naturalness of synthetic speech.更多还原