Articulatory Speech Synthesis for Natural User Interfaces

多年来，发音合成的研究在很大程度上被基于共振峰和基于声学的语音合成技术所掩盖。虽然在某些领域(例如，基于语音的数据库)取得了成功，但这些技术仍然不能从任意说话者的文本中产生自然的外观和发音的语音。自然的外观和声音的语音技术是自然用户界面的语音交互的下一个主要里程碑之一。发音语音合成在工业界和学术界都取得了稳步的进展，现在已经准备好为克服语音产生中的基本问题提供必要的平台，我们相信这代表着语音合成技术的下一个重大进步。 由于人类声道和语音产生行为的结构复杂性，以前的三维发音合成研究主要集中在对语音产生和声道结构的狭义方面进行分析和建模。不是对整个声道的几个子组件进行建模以产生有限的非自然发声集合，而是需要一个更完整的平台，该平台将允许在利用用于整个声道的最佳可用技术的工作的发音语音合成器的环境中集成和测试声道子组件。 几十年来，哈斯金斯2D节律语音合成器一直被广泛使用，尽管它可以产生的形状和声音存在众所周知的局限性，并且缺乏通用或特定于说话者的生产参数的准确表示。诸如Birkholtz的VTL等进展已经在3D发音合成方面取得了进展，但在视觉上仍然不发达，并且缺乏生物力学基础。为了克服这些局限性，并为发音语音合成的新研究提供一个平台，我们提出了一种空气动力学驱动的发音语音合成器，该发音合成器基于一个全面的、参数化的声音和面部发音器的三维生物力学模型，能够产生可视和有声语音和非语音。

For years, articulatory synthesis research has been largely overshadowed by formant-based and acoustic-based speech synthesis techniques. While successful in some domains (e.g., voice-based databases), these techniques still cannot produce natural looking and sounding speech from text from an arbitrary speaker. Natural looking and sounding speech technology is one of the next major milestones in voice-based interaction for natural user interfaces. Articulatory speech synthesis has progressed steadily at the fringes of both industrial and academic interest and is now poised to provide the necessary platform to overcome basic problems in speech production and, we believe, represents the next major advance in speech synthesis technology. Because of the structural complexity of the human vocal tract and of speech production behaviour, prior research in 3-dimensional articulatory synthesis has been focused on analyzing and modeling narrowly defined aspects of speech production and vocal tract structure. Rather than modeling a few sub-components of the overall vocal tract for production of a limited set of unnatural utterances, a more complete platform is needed that will allow vocal tract sub-components to be integrated and tested within the context of a working articulatory speech synthesizer that utilizes the best available technologies for the entire vocal tract. For decades, the Haskins 2D Articulatory Speech Synthesizer has been commonly used, even with the well-known limitations of shapes and sounds it can produce, and the lack of accurate representations of either generic or speaker-specific production parameters. Advances, such as VTL by Birkholtz, have made progress in 3D articulatory speech synthesis, but remain visually undeveloped as well as lacking biomechanical foundations. To overcome these limitations and provide a platform for new research in articulatory speech synthesis, we propose to construct and evaluate an aerodynamically driven articulatory speech synthesizer based on a comprehensive, parameterized 3D biomechanical model of the vocal and facial articulators, that is capable of producing both visible and acoustic speech and non-speech.