Determinants of perceptual learning for speech perception

认知科学中最大的谜团之一是，尽管所产生的声音的实际声学差异极大，但人类如何理解语言。说话者很少发出相同语音的相同版本，因为每个人都有自己的说话方式，这取决于他们的方言、性别和年龄。其他因素，如疲劳或兴奋，可能会随时间而变化。尽管听者通常能够适应这些差异，但研究尚未发现这种适应到底是如何起作用的。这个项目建立了一种新的工具来创造实验刺激，这种刺激可以产生自然语言中发生的广泛变化，并使用这个工具来检查听者如何适应说话者的差异。这些实验测试了听者如何从说话者的声音中吸收新的证据，他们如何使用关于自己语言中语音的现有知识，他们如何学会忽略什么，以及个体在适应说话者方面的能力有何不同。该项目具有潜在的广泛社会效益，因为它将提供人们如何相互交流的基础知识，并为本科生和研究生提供实践教育，他们将发展实验设计、计算机编程、数据分析和科学交流的技能。该项目还将为其他研究人员创造工具，包括开放获取的刺激语料库和简短的讲座视频和作业，可用于向学生和其他人传授声物理学，从而促进公众的科学素养。这个跨学科的项目结合了心理语言学实验、信号处理和计算模型来测试语音适应的贝叶斯信念更新模型的三个关键部分：(1)信念反映了听话者对线索分布和语音类别之间关系的知识，(2)适应反映了观察到的证据与先前信念的整合，这种整合可以由无监督和有监督的学习信号驱动，(3)信念更新是特定于语境的(例如，以说话者为条件)。该项目将建立一种创建摩擦连续体的新方法，创建一个广泛的刺激语料库，并比较刺激语料库和自然刺激之间的学习。这种新颖的刺激产生方法将被用来产生声音，这些声音独立地操纵频率缩放(反映声道大小)并获得摩擦线索的属性(反映特定的语音清晰度)，以发现哪些特定的线索指导学习，以及这些信息如何在说话者之间概括(或不概括)。这些实验还将考察听者如何随着时间的推移整合无监督和词汇监督的学习信号，从而洞察先前暴露对词汇指导学习的影响以及学习中的个体差异。这些都是重要的创新，因为它们解决了现有知觉学习文献中的两个弱点，包括叠加两个声音以产生可能不自然的模棱两可的输入的做法，以及未能衡量感知随时间的增量变化。理论上的好处包括对言语适应的贝叶斯信念更新模型进行了强有力的测试，确定了驱动知觉学习的特定线索，以及揭示了知觉学习如何整合无监督和词汇监督的信号。该奖项反映了NSF的法定使命，并通过使用基金会的智力价值和更广泛的影响审查标准进行评估，被认为值得支持。

One of the great mysteries in the cognitive sciences is how humans understand speech despite extreme variation in the actual acoustics of the sounds produced. Speakers rarely produce identical versions of the same speech sound, because everyone has a personal way of talking depending on their dialect, gender, and age. Other factors, such as fatigue or excitement, might change moment-to-moment. Although listeners are usually able to adapt to these differences, research has yet to discover exactly how this adaptation works. This project establishes a new tool for creating experimental stimuli that can generate the wide variation that occurs in natural speech and uses this tool to examine how listeners adapt to talker differences. The experiments test the ways that listeners incorporate new evidence from a talker’s voice, how they use existing knowledge about the speech sounds in their language, how they learn what to ignore, and how individuals may differ in their ability to adapt to a talker. The project has potentially broad societal benefits because it will provide foundational knowledge on how people communicate with each other and provide hands-on education for undergraduate and graduate trainees who will develop skills for experimental design, computer programming, data analysis, and science communication. The project will also create tools for other researchers, including an open-access stimulus corpus and brief lecture videos and assignments that can be used to teach students and others about the physics of sound, thus promoting scientific literacy among the public. All materials associated with this project will be made publicly available.This interdisciplinary project unites psycholinguistic experimentation, signal processing, and computational modeling to test three key parts of a Bayesian belief-updating model of speech adaptation: (1) beliefs reflect listeners’ knowledge of the relationship between cue distributions and phonetic categories, (2) adaptation reflects integration of observed evidence with prior beliefs that can be driven by both unsupervised and supervised learning signals, and (3) belief-updating is context-specific (e.g., conditioned on talker). The project will establish a novel method for creating fricative continua, create an extensive stimulus corpus, and compare learning between the stimulus corpus and natural stimuli. The novel method of stimulus creation will be used to generate sounds that independently manipulate frequency scaling (reflecting vocal tract size) and gain properties (reflecting specific speech articulations) of fricative cues to discover which specific cues guide learning and how that information generalizes (or not) across talkers. The experiments will also examine how listeners integrate unsupervised and lexically supervised learning signals over time, providing insight into the influence of prior exposure on lexically guided learning and individual differences in learning. These are important innovations because they address two weaknesses in the existing perceptual learning literature, including the practice of superimposing two sounds to create ambiguous input that might be unnatural and the failure to measure incremental change in perception over time. Theoretical benefits include a strong test of the Bayesian belief-updating model of speech adaptation, identification of specific cues that drive perceptual learning, and unpacking how perceptual learning integrates unsupervised and lexically supervised signals.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.