摘要
基频曲线预测是文语转换系统中韵律控制的重要内容。基于决策树的分析,本文提出了一个采用三个控制参数,修改一组归一化的音节基频曲线模板,从而生成连续语流基频曲线的预测模型。由于前一音节的声调动态目标将在该音节的偏后部分实现,甚至影响到连接紧密的后一个音节的起始部分,在预测当前音节的控制参数时,前一音节的基频终止值也作为语境参数之一,参与预测,其结果既保持了前后音节基频的连续性,也提高了基频曲线整体预测的准确率。测试表明:预测基频与实际基频的音节内标准误差小于10 Hz。将这一方法应用于PSOLA语音合成系统后,合成语音的自然度令人满意。
Based on decision tree analysis, this paper proposes an approach of F0 contours prediction for natural speech, in this approach, syllabic maximum pitch, minimum pitch and duration, as well as a set of normalized template of syllabic F0 shape are used to generate F0 contours for each syllable. To maintain the pitch continuity on sentence level, the predicted pitch information of previous syllable is combined with context parameters to predict the pitch contour of current syllable. In testing data, the syllabic RMS error between real pitch and predicted pitch is less than 10 Hz. This fact demonstrates the continuity of pitch contours and high prediction accuracy. Finally, it achieves natural sounding prosody when applying this method to a PSOLA speech synthesis.
出处
《声学学报》
EI
CSCD
北大核心
2006年第1期19-27,共9页
Acta Acustica
