Collaborative Research: A Bayesian model of phonetic and phonotactic effects in cross-language speech production

人类认知的一个基本方面是感知和产生语言的能力。对非母语语音处理的研究提供了一些与这种能力有关的最显著的证据：当人类试图感知或产生包含外来声音或声音序列的单词时，他们会表现出正确和错误的系统性模式。以前的研究已经证实，不同的非母语结构会导致不同的错误率和类型；有人假设，这些差异可以由语法、感知和发音等因素共同解释。这个项目的主要目标是对这些因素提供仔细受控的实验评估，并开发一个明确的、概率的模型，说明它们如何在人类表现中相互作用。要研究的具体实验问题是：(1)人类在处理非母语声音时对哪些语音特征最敏感；(2)输入和环境声学的质量如何影响非母语语音的感知和产生；以及(3)学习词义是否可以调节对非母语语音细节属性的敏感性。计算模型建立在越来越多的工作基础上，表明人类的感知和行动反映了最优的贝叶斯推理，条件是先前的预期和嘈杂的感觉测量。相关的先验知识反映了对母语的了解；该模型预测，与母语中的结构更相似的非母语结构应该被更准确地处理。该模型还预测，具有稳健感知特性的非原生声音应该被更准确地处理，即使它们的先验概率很低。这一模型的开发将提供给其他研究人员，将在认知科学研究的更广泛背景下促进音系学和语音学的作用。由于我们的实验考察了课堂声学和说话者变化对非母语声音处理的影响，该项目也对外语教育学产生了影响。

A fundamental aspect of human cognition is the capacity to perceive and produce language. Studies of non-native speech processing provide some of the most striking evidence bearing on this capacity: when humans attempt to perceive or produce words containing foreign sounds or sound sequences, they show systematic patterns of correct and incorrect performance. Prior research has established that different non-native structures elicit different rates and types of error; it has been hypothesized that these differences can be explained by a combination of grammatical, perceptual, and articulatory factors. The main goals of this project are to provide carefully controlled experimental evaluations of these factors, and to develop an explicit, probabilistic model of how they interact in human performance. Particular experimental issues to be investigated are: (1) what phonetic characteristics humans are most sensitive to when processing non-native sounds; (2) how the quality of the input and ambient acoustics affect non-native perception and production; and (3) whether learning word meaning can modulate sensitivity to detailed properties of non-native sounds. The computational model builds on a growing body of work suggesting that human perception and action reflect optimal Bayesian inference conditioned on prior expectations and noisy sensory measurements. The relevant prior reflects knowledge of the native language; the model predicts that non-native structures that are more similar to those in the native language should be processed with greater accuracy. The model also predicts that non-native sounds with robust perceptual properties should be processed more accurately, even if their prior probabilities are low. The development of this model, which will be made available to other researchers, will promote the role of phonology and phonetics within the broader context of cognitive science research. Because our experiments examine the impact of classroom acoustics and talker variation on non-native sound processing, this project also has ramifications for foreign language pedagogy.