论文部分内容阅读
正确标记短语间的停顿,对提高文语转换系统合成语音的自然度起着重要作用。介绍一种采用最大熵模型从真实自然的语音流中自动识别汉语短语间停顿的方法。模型的特征集包含语音和词法两类特征,采用半自动的方式获得。首先由人工根据经验设计候选特征集,然后采用特征选择算法对候选特征进行筛选,选择更有效的特征构成最终特征集,并训练生成用于汉语短语间停顿识别的最大熵模型。3组实验的结果表明,模型能够取得比较满意的短语间停顿识别效果。
The pauses between correctly labeled phrases play an important role in improving the naturalness of synthetic speech in the text-to-speech conversion system. A method of automatically identifying the pauses between Chinese phrases from real natural speech streams using the maximum entropy model is introduced. The model’s feature set contains two types of speech and lexical features, obtained in a semi-automatic way. Firstly, the candidate feature set is designed by human based on experience. Then the feature selection algorithm is used to filter the candidate features, the more efficient features are selected to form the final feature set, and the maximum entropy model for the pause recognition between Chinese phrases is generated. The results of the three groups of experiments show that the model can achieve a satisfactory recognition effect of pause between phrases.