ITR: Evaluation and Personalization of Synthetic Voices

摘要本项目解决了计算机生成自然声音语音输出的两个关键问题。第一个方面提供了合成语音的评价策略。由于没有明确的客观指标来衡量语音合成质量，很难显示系统的改进和比较技术。这项工作提供了一份详细的评估技术清单，包括合成过程的子部分：文本分析，词汇发音，韵律和语音质量。这些技术的目的是对其他语言中现有的和新开发的声音起作用，在这些语言中，衡量标准往往更难获得。适当的人体实验是用来证明给定的指标。这项工作的第二个方面是改进说话人特定语音的建模。在单元选择合成技术中，将词汇与特定说话人的实际说话形式相匹配可能会在选择过程中引入显著的噪声。说话者的方言和风格会有特定的方言。使用数据驱动技术，在基于知识的框架内，我们为单元选择合成导出适当的分割类型。结果通过CMU的http://festvox.org网站发布。

Evaluation and Personalization of Synthetic VoicesAbstractThis project addresses two key issues in the generation of natural sounding spoken output by computer. The first aspect provides an evaluation strategy for synthetic voices. With no clear objective measure for speech synthesis quality, it is hard to show improvement in systems and compare techniques. This work provides a detailed list of evaluation techniques for sub-parts of the synthesis process including: text analysis, lexical pronunciation, prosody, and phonetic quality. The techniques are designed to act on existing and newly developed voices in other languages, where measures are often harder to come by. Appropriate human experimentation is used to justify the given metrics.The second aspect of this work is to improve modeling of speaker-specific acoustic phonetics. In unit selection synthesis techniques, matching the lexicon with the actual spoken form of a particular speaker can introduce significant noise in the selection process. Speakers will have particular idiolects for their dialect and style. Using data-driven techniques, within knowledge-based frameworks, we derive appropriate segmentation types for unit selection synthesis.Results are released through CMU's http://festvox.org site.