CAREER: Understanding Real-World Auditory Scene Analysis

听觉科学中的一个基本问题是人们如何在竞争声源的存在下识别语音和其他声音，就像在拥挤的餐馆里与晚餐伙伴交谈一样。当感兴趣的声音嵌入其他声音的混合物中时，听到感兴趣的声音的过程被称为“声音分离”，并且人类听众在分离声音方面远远优于机器系统。然而，这一过程往往是费力的，极易受到听力损伤的影响，包括通常伴随正常衰老的听力损伤。了解人类听者声音分离的基础，以及限制人类分离能力的因素，将有助于开发用于鲁棒语音识别和声音识别的辅助听力设备和机器系统。该项目还将通过一系列公开的在线视频演示来激发公众和中学生对听力的兴趣。该CAREER奖项旨在通过探索声音与自然声音分离的基础来丰富对人类听觉感知的理解。这些实验将利用语音分析和合成方法的最新进展，1）操纵自然语音中的分组线索，并测试它们对人类听众声音分离的影响; 2）操纵语音和语音结构，以探索它们在分离中的作用; 3）测试人类听众关注和跟踪目标声源的能力。长期目标是激发信号处理算法，促进人类听众的隔离，并在机器系统中复制他们的能力。

A fundamental question in auditory science concerns how people can recognize speech and other sounds in the presence of competing sound sources, as when conversing with a dinner partner at a crowded restaurant. The process of hearing a sound of interest when it is embedded in a mixture of other sounds is known as "sound segregation" and human listeners vastly outperform machine systems for segregating sounds. However, the process is frequently effortful, is highly vulnerable to hearing impairment, including hearing impairment that typically accompanies normal aging. Understanding the basis of sound segregation in human listeners, and the factors that limit human segregation abilities, would enhance efforts to develop assistive listening devices and machine systems for robust speech recognition and sound recognition. The project will be complemented by an educational effort to stimulate interest in audition in the general public and in middle- and high-school students through a series of publicly available online video presentations describing auditory research with associated sound demonstrations.This CAREER award is aimed at enriching the understanding of human auditory perception by exploring the basis of sound segregation with natural sounds. The experiments will leverage recent advances in speech analysis and synthesis methods to 1) manipulate grouping cues in natural speech and test their effect on sound segregation in human listeners; 2) manipulate voice and speech structure to probe their role in segregation; and 3) test the ability of human listeners to attend to and track target sound sources. The long-term goals are to inspire signal-processing algorithms that facilitate segregation by human listeners and replicate their competence in machine systems.