论文部分内容阅读
近年来提出的基于隐Markov模型的单元挑选语音合成方法,较好地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,但该方法在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。该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。
In recent years, the unit selection speech synthesis method based on Hidden Markov model, which solves the problem of dependence on more manual intervention and instability of synthesis in the traditional splicing synthesis, solves the problem of synthesizing different acoustic statistical models When using the model weights can not be automatically trained to obtain, and artificial optimization more difficult. This paper presents a model weight optimization method based on synthetic quality prediction. The method first collects less artificial audiometry results and constructs a prediction model for synthetic voice quality under different weights using multivariate adaptive regression splines. Then, the optimal weight value is automatically searched based on the predictive model using a pattern search algorithm. Experimental results show that this method can effectively optimize the model weights and improve the naturalness of synthesized speech.