Naturally Sounding Speech Synthesis and Recognition Based on the Formulation of Prosody

本研究旨在阐明语音韵律特征与语言信息、准语言信息和非语言信息之间的关系，实现语音合成的先进技术：1.通过对基频轮廓进行低通滤波来抑制重音成分并提取其偏差值，从而提高了从基频轮廓中自动提取短语成分的精度。在自动韵律标注方法中实现了准确率的进一步提高，该方法利用从语言信息中获得的韵律知识作为约束F0参数估计。构造了类对话语音合成的Mora时长规则。这些规则基本上是将阅读风格的每个Mora时长修改为以韵律短语为基础的对话式演讲的时长，由FO轮廓定义。不同态度…语音的韵律特征更多的是对S/情绪的分析。研究发现，说话者有选择地控制几个韵律线索来表达态度/情感的程度。通过知觉实验还发现，分段特征控制对于实现情感话语也是不可或缺的。提出了一种以Mora为单位用代码表示韵律词的F0轮廓并对其转换进行统计建模的方法(Moraic转换的统计模型)。韵律词边界的插入误差为11%~15%，检测率为70%~75%。将该方法应用于连续语音识别，对Mora识别率的改善不大。提出了一种利用重音类型和短语边界位置的输入生成句子F0轮廓的方法。针对大词汇量连续语音识别波束搜索过程中的动态剪枝问题，提出了一种基于韵律特征的动态剪枝方法。实验证明，在不降低识别率的情况下，搜索空间可以减少到四分之一。该方法增大了韵律边界处的波束宽度，减小了边界之间的波束宽度。提出了一种利用韵律边界信息选择具有不同语境依赖关系的音素模型的方法。在此基础上，开发了一个学术信息检索口语对话系统，并对该系统进行了评价。较少

Several results including the following ones were achieved through the study aiming at formulating the relationship between prosodic features of speech and linguistic and para/non linguistic information, and realizing advanced technologies on speech synthesis :1. An improved accuracy was realized in automatic extraction of phrase component onsets from fundamental frequency (FO) contours by suppressing accent components through low-pass filtering of the contours and by taking their deviations. Further improvements in accuracy were realized in a method of automatic prosodic labeling where knowledge on prosody obtainable from linguistic information was utilized as constrictions F0 parameter estimation.2. Mora duration rules were constructed for dialogue-like speech synthesis. These rules are basically to modify each mora duration of reading-style speech to that of dialogue-like speech in prosodic phrase-basis, defined by the FO contours.3. Prosodic features of Speech with various attitude … More s/emotions were analyzed. It was found that a speaker selectively controlling several prosodic cues to express degree of attitudes/emotion. It was also found through a perceptual experiment that segmental feature control were also indispensable to realized emotional speech.4. A method was developed to represent F0 contours of prosodic words by codes in mora unit and to model their transitions statistically (Statistic model of moraic transition). The detection rates of 70-75% were achieved with insertion errors of 11-15% for prosodic word boundaries. The method was applied to continuous speech recognition with few % improvements in mora recognition rates. A method was also developed to generate sentence F0 contours with inputs of accent types and phrase boundary positions.5. A prosodic feature-based method was developed for the dynamic pruning in beam search process of large-vocabulary continuous speech recognition. It was proved that the search space could be reduced to a quarter without degradation in recognition rates. The method enlarges beam width at prosodic boundaries and decreases between boundaries. A method was also developed to select phoneme models with various context dependencies using prosodic boundary information.6. Based on the results obtained, a spoken dialogue system of academic information retrieval was developed and evaluated. Less