论文部分内容阅读
提出了一种基于决策树的语音合成基元的语境特征权重训练算法.对语音数据库中的每个带调音节,利用语境相关的问题集和候选基元的频谱距离建立决策树.对每个要合成的音节,根据其语境特征,获得语音合成系统选择的基元的语境特征F*和该语境特征下决策树叶子结点中基元的语境特征F′.统计F′中每一个语境特征相对于F*的变化,根据语境特征变化的概率对权重进行调整.实验结果表明,这种方法能够训练出合理的语境特征权重,使得合成语音的自然度有一定提高.同时,利用这种方法还可以对语音合成系统进行实时优化.
A context-based feature weight training algorithm for speech synthesis primitives based on decision tree is proposed in this paper. For each syllable in the speech database, a decision tree is constructed using the context-dependent set of questions and the spectral distances of candidate primitives. For each syllable to be synthesized, the contextual features F * of the primitives selected by the speech synthesis system and the contextual features F ’of the primitives in the leaf nodes of the decision tree under the contextual features are obtained according to their contextual features. Statistics F ’, We adjust the weight of each of the contextual features relative to F * according to the probability of the change of contextual features.The experimental results show that this method can train the reasonable weight of contextual features and make the naturalness of the synthesized speech Must be improved at the same time, the use of this method can also be real-time voice synthesis system optimization.