基于数据挖掘的普通话韵律规则学习

Data Mining for Learning Mandarion Prosodic Models

普通话韵律规则对于语音合成和语音学研究具有重要意义 .为了更有效地进行韵律规则学习 ,该文利用数据挖掘技术从语料库中提取规则 .通过聚类分析进行基频模式提取 ,并以此进行基频序列的离散化 ;由语言学分析的结果得出训练句子中每个音节的参数 ,利用决策树和神经网络学习音节的韵律变化规则 .测试表明基于数据挖掘的韵律规则学习取得了较好的结果 ,证实了方法的有效性 .

Mandarin prosodic models are very important in speech research and speech synthe sis, which mainly describess the variation of pitch. The models that are now being u sed in most Chinese Text\|To\|Speech systems are constructed by expert, qualitatively an d with low precision. In this paper, Data Mining is used to extract more accurate prosodic pattern s from actual large mandarin speech database to improve the naturalness and intelligibility of synth esized speech. In data preprocessing, typical prosody models are found by clustering analysis, a nd the original pitches extracted from sentences are discrete with classic pitch models. These clusters together with some linguistic features (including tone combination, word length, part\|of \|speech (POS), syllable position in word, word position in phrase) obtained by text parsing are use to acquire training data. ANN and Decision tree are trained respectively using above integr ated data to learn the variation prosody models of pitch. Two decisino trees are construc ted for predicting the classic pitch model and length of pitch based on C4.5, and BackPropagation(BP) network is used to learn the mapping between the linguistic features and the mean value of pit ch. Encouraging experimental results show the effectiveness of the proposed method base on Data Mining.