固有声(eigenvoice)に基づいた音声合成---多様な声質の実現を目指して---

在这项研究工作中，我们研究了一种基于HMM的语音合成系统的特征语音技术，该系统可以合成不同音质的语音。在基于HMM的语音识别中用于快速说话人自适应的特征语音技术中，大量依赖于说话人的HMM集合通过降维技术(例如，PCA)由几个参数来表示。需要适应的参数很少，并讨论了这些参数与语音质量之间的对应关系。因此，我们推测，通过将特征语音技术应用于基于HMM的语音合成系统，我们可以很容易地指定语音质量并合成所需的语音。基于这种观点，我们提出了一种用于语音合成的特征语音技术，并将其应用到基于HMM的语音合成系统中，该系统在HMM的统一框架内同时对频谱和FO进行建模。因此，用户可以通过设置几个参数来定义语音质量：特征语音的权重。我们进行了主观评价测试，以调查特征语音的权重值与语音质量之间的对应关系。根据主观评价测试的结果构建的语音合成系统可以合成各种音质的语音，通过设置代表音质的参数，我们证实了很容易获得期望的音质的可能性。

In this research work, we investigate an eigenvoice technique for an HMM-based speech synthesis system which can synthesize speech with various voice qualities. In the eigenvoice technique for very fast speaker adaptation in HMM-based speech recognition, a large number of speaker dependent HMM sets are represented by a few parameters through a dimensionality reduction technique, e.g., PCA. The parameters to be adapted are very few, and the correspondence between the parameters and voice qualities was discussed. Accordingly, we surmise that by applying the eigenvoice technique to the HMM-based speech synthesis system, we can specify voice qualities and synthesize desired voices easily. In this point of view, we proposed an eigenvoice technique for speech synthesis and applied it to the HMM-based speech synthesis system, which models spectrum and FO simultaneously in a unified framework of HMM. As a result, users can define voice qualities by setting a few parameters : weights for the eigenvoices. We conducted subjective evaluation tests to investigate the correspondence between weight values for eigenvoices and voice qualities. The speech synthesis system constructed based the results of the subjective evaluation tests can synthesize speech with various voice qualities and we confirmed the possibility of obtaining desired voice qualities easily by setting the parameters which represent the voice qualities.