Synthesis of speech in any speaking styles based on corpus-based generation of prosodic features using the generation process model

研究建立了一种基于语料库的语音合成方法，该方法基于基频轮廓生成过程模型，可以在任何说话风格下生成高质量的语音。最初的研究计划得到了以下结果：1。提出了一种利用二叉决策树对生成过程模型的命令参数进行预测的方法，该方法可通过文本解析获得语言信息等输入，从而合成基频轮廓。将上述方法与其他方法相结合，利用二叉决策树预测停顿位置、停顿长度和音素持续时间，实现了韵律控制的集成方法。通过包括情绪语音在内的多种风格语音合成实验，验证了该方法的有效性。提出了一种利用二值决策树从观测到的基频轮廓中自动提取命令参数的方法。结果表明，在树的输入中加入文本的语言信息可以提高提取的准确性。构建二叉决策树来预测具有特定焦点的话语在短语和重音命令上的偏差。它们的输入是聚焦词在句子中的重音类型和位置，以及没有特定聚焦的话语中对应部分的命令值。根据预测偏差，通过修改第1节方法预测的短语和重音命令来实现适当的焦点控制。提出了一种两步法生成标准汉语基频轮廓的方法。它首先以基于语料库的方式生成短语成分，然后以基于语料库的方式生成语气成分。该方法在合成基频轮廓时具有较高的灵活性。作为柔性控制的一个实例，表明用一组简单的规则可以实现适当的对焦控制。将上文第1节和第2节开发的方法与HMM语音合成相结合，构建了日语和汉语语音合成系统。结果表明，与在HMM框架下进行韵律控制的“全”HMM合成器相比，该系统可以实现更高自然度的合成语音。实验结果表明，该系统可实现多种风格的合成语音。利用上述语音合成系统构建了道路引导和电视节目引导语音对话系统。通过根据用户的性格和情境控制回复语音的说话风格的实验，证明了所开发的语音合成方法的有效性。少

Research works were conducted to establish a corpus-based speech synthesis method, which is based on generation process model of fundamental frequency contours and can generate high-quality speech in any speaking styles. The original research plan was fulfilled with the following results :1. A method was developed to predict the command parameters of the generation process model using binary decision trees with inputs such as linguistic information available by parsing texts, and thus to synthesize fundamental frequency contours. An integrated method of prosodic control was realized by integrating the above method with other methods using binary decision trees to predict pause positions and lengths and phoneme durations. The validity of the method was shown through experiments on speech synthesis of various styles including emotional speech. A method was also developed to automatically extract the command parameters from observed fundamental frequency contours using binary decision tre … More es. It was shown that the accuracy of extraction increased by including linguistic information of the text into inputs of the trees.2. Binary decision trees were constructed to predict deviations in phrase and accent commands of the utterances with specific focuses from those without. Their inputs are accent types and positions in sentences of the focused words, and command values of the corresponding parts of the utterances without specific focus. An appropriate focus control was realized by modifying the phrase and accent commands predicted by the method in section 1 based on the predicted deviations.3. A two-step method was developed for generating fundamental frequency contours of Standard Chinese. It first generates phrase components in a corpus-based way, and then generates tone components in a corpus-based way. The method has a high flexibility in synthesizing fundamental frequency contours. As an example of flexible control, it was shown that proper focus control could be realized in a simple set of rules.4. Speech synthesis systems were constructed for Japanese and Chinese by integrating methods developed in sections 1 and 2 above with HMM speech synthesis. It was shown that synthetic speech with higher natural ness could be realized by our system than using "full" HMM synthesizer, where prosodic control was done in HMM framework. It was also shown that various styles of synthetic speech could be realized by our system.5. Spoken dialogue systems for road guidance and TV program guidance were constructed using the above speech synthesis systems. The validity of the developed speech synthesis method was proved through experiments on the control of speaking styles of reply speech depending on the user's characters and situations. Less