Investigating neural mechanisms for flexible, robust speech perception with fMRI

描述（由申请人提供）：大脑受到来自世界的感官信息的轰炸，必须使用有限的神经资源提取某些有用的信息。这意味着大脑必须是高效的，扔掉不需要的信息，以便专注于来自世界的感官信息中最重要的部分。然而，在一种情况下没有信息的信息在另一种情况下可能是非常有用的，因此这种效率必须与灵活性相匹配。这一点在言语感知领域尤为明显，在言语感知中，嘈杂、模糊的感觉信号被映射到潜在的语言单位，如音素、单词和句子。这种映射会根据谁在说话而发生很大变化。大脑可能处理这个问题的一种方法是学习说话者特定的表征，优化语音处理的效率，并在说话者改变时部署或“交换”这些表征，必要时为新的说话者学习新的表征。虽然有一些证据表明听众确实使用这种策略，但对潜在的神经机制知之甚少。这项建议旨在通过两个具体目标澄清这些机制。首先，功能性磁共振成像（fMRI）将对听者的大脑进行成像，同时他们正在听两个不同口音的说话者说话，混合在一起。通过比较说话者切换时活跃的区域与学习每种口音期间活跃的区域（如通过行为反应测量的），将阐明听者学习和部署说话者特定表征的回路。第二，使用多体素模式分析技术，相同的语音的神经表征，具有不同的解释取决于说话者将被测量，以确定如何深入说话者的特定知识影响语音的处理。如果说话者的特定知识被用来优化低水平感知处理的效率，那么类内差异应该导致更相似的活动模式，而跨类差异应该导致更不同的活动模式。

DESCRIPTION (provided by applicant): The brain is bombarded by sensory information from the world, and must extract certain pieces of useful information using limited neural resources. This means that the brain must be efficient, throwing away information that is not needed in order to focus on the most important part of the sensory information from the world. However, information that is uninformative in one situation may be highly informative in another, and thus this efficiency must be matched by flexibility. One domain where this is particularly true is speech perception, where a noisy, ambiguous sensory signal is mapped onto underlying linguistic units like phonemes, words, and sentences. This mapping changes substantially depending on who is talking. One way the brain might deal with this is to learn talker-specific representations which optimize the efficiency with which speech sounds are processed, and deploy or "swap out" those representations whenever the talker changes, learning new representations for new talkers as necessary. While there is some evidence that listeners do use such a strategy, little is known about the underlying neural mechanisms. This proposal seeks to clarify these mechanisms through two specific aims. First, functional magnetic resonance imaging (fMRI) will image the brains of listeners while they are hearing words from two talkers with different accents, mixed together. By comparing the areas that are active when the talker switches with areas that are active during periods of learning about each accent (as measured by behavioral responses), the circuits by which listeners learn and deploy talker-specific representations will be elucidated. Second, using multi-voxel pattern analysis techniques, the neural representations of identical speech sounds which have different interpretations depending on the talker will be measured to determine how deeply talker-specific knowledge affects the processing of speech sounds. If talker-specific knowledge is being used to optimize the efficiency of perceptual processing at a low level, then within-category differences should result in more similar patterns of activity, while across-category differences should result in more distinct patterns of activity.