Doctoral Dissertation Research: Predictability and Recoverability of High Vowel Reduction in Speech

人类认知依赖于直接和间接的信息来得出结论。依赖直接信息的过程称为自下而上的过程，其中仅根据提供的信息得出结论；依赖间接信息的过程称为自上而下的过程，其中使用上下文、模式识别和概率等元素来形成对特定结果的预期。研究表明，强大的自上而下的过程往往会阻碍对刺激中某些信息的自下而上的处理，但这两种类型的过程是如何相互作用的，人们知之甚少。作为人类认知不可分割的一部分，语言也表现出自上而下和自下而上的相互作用，在一种语言中有意义的差异在另一种语言中很容易被忽视，因为它没有提供信息。因此，语言是一个很好的工具，可以揭示大脑中直接和间接信息的处理是如何相互作用的。日语的音节结构绝大多数遵循简单的辅音-元音结构。然而，在日语中，在两个无声辅音(如/suki/[ski]‘like’)之间弱化不重读的元音/i、u/的过程中，往往会导致在语音产生过程中完全删除一个元音。然而，说日语的人经常报告说，即使在完全没有元音的情况下，也能感觉到两个连续的清音辅音之间有元音(例如，[ski]被认为是/suki/)。目前还不清楚是什么导致了这种对明显不存在的元音的误解，但可能有两个来源。第一个是间接来源。由于日语的音节结构简单，说日语的人习惯于在辅音之后加上元音，在这种情况下，对元音的误解将是一个强烈的自上而下过程的结果。另一种选择是直接来源。由于高元音减少在日语中很常见，日语使用者甚至可能对指示元音存在的细微声学信息高度敏感，在这种情况下，元音的感知将是自下而上过程的结果。一项发音实验和一项语料库研究将首先测试自然产生的语音中存在哪些针对弱化元音的声学信息。然后，感知实验将操纵直接和间接信息，以测试这些日语听者在感知到元音减弱时更依赖于哪一个。实验结果将提供给一个计算模型，该模型调查对直接和间接线索的依赖可能是如何学习的。这些实验和这个模型一起，将使我们更深入地理解自下而上和自上而下的过程在多大程度上受到语言经验的影响，以及更广泛地说，在人类整体认知中。

Human cognition relies on direct and indirect information to draw conclusions. Processes that rely on direct information are called bottom-up processes, where a conclusion is drawn based solely on the provided information; processes that rely on indirect information are called top-down processes, where elements such as context, pattern-recognition, and probability are used to form an expectation for a particular outcome. Studies have shown that a strong top-down process often blocks bottom-up processing of certain information available in the stimulus, but how the two types of processes interact is poorly understood. As an integral part of human cognition, language also shows interactions of top-down and bottom-up processes, where differences that are meaningful in one language may be readily ignored as uninformative in another. Language, therefore, is an excellent tool for shedding light on how the processing of direct and indirect information interacts in the mind.This research focuses on Japanese as a case study. The syllable structure of Japanese overwhelmingly follows a simple consonant-vowel structure. However, the process of high vowel reduction in Japanese, where unaccented vowels /i, u/ are weakened between two voiceless consonants (e.g., /suki/ [ski] 'like'), often results in the complete deletion of a vowel during speech production. However Japanese speakers often report perceiving a vowel between two consecutive voiceless consonants, even in the complete absence of a vowel (e.g., [ski] perceived as /suki/). It is unclear at this point what drives this misperception of apparently non-existent vowels, but there are two likely sources. The first is an indirect source. Given the simple syllable structure of Japanese, Japanese speakers are conditioned to expect a vowel after a consonant, in which case the misperception of a vowel would be the result of a strong top-down process. The alternative is a direct source. Because high vowel reduction is common in Japanese, Japanese speakers may be hypersensitive even to subtle acoustic information that indicates the presence of a vowel, in which case the perception of a vowel would be the result of a bottom-up process. A production experiment and a corpus study will first test what acoustic information for a weakened vowel exists in naturally produced speech. A perception experiment will then manipulate the direct and indirect information to test which of these Japanese listeners rely on more when perceiving a weakened vowel. The experimental results will inform a computational model, which investigates how a reliance on the direct and indirect cues might be learned. The experiments, together with the model, will lead to a deeper understanding of the extent in which bottom-up and top-down processes are shaped by experience both in language and more broadly in human cognition as a whole.