AMC-SS: Dynamic Algorithms For Blind Separation Of Convolutive Sound Mixtures

主要研究人员和同事将研究卷积声音混合的盲源分离算法的数学和计算问题。当有多个扬声器时，在任何封闭的环境中都会出现卷积混合声音。目标是基于源独立性和在多个位置接收的数据，在没有环境先验知识的情况下，自适应地实现混合物的动态分离和源恢复。虽然快速傅立叶变换有助于将问题定位在频域中，并且现有方法在每个频率上的分离都相当成功，但排列和缩放问题仍然是不确定的，可能会极大地影响分离质量。研究人员将使用动态更新的统计信号信息来修复排列，并在时间域中最小化混合滤波器长度来修复缩放。他们还将研究时间域中的动态算法，通过优化混合过滤器长度和源独立性，以及利用概率理论来研究恢复的混合过滤器的动态稳定性和收敛。该项目旨在开发分离真实声音混合的算法，这是机器(计算机)无法像人类一样出色地完成的任务。这一挑战也被称为鸡尾酒会问题，是提高现代听力设备质量的根本。例如，众所周知，助听器和耳蜗植入物在安静的情况下效果很好，但当有竞争的声源时，它们会迅速退化。即使对于听力正常的人来说，通过来自相当嘈杂的地方(如餐馆)的手机通话也很难进行对话。理解盲源分离的数学原理并将其应用于实际计算是解决这一问题的关键步骤。该项目在对信息技术和生物技术的进步产生广泛影响和作出数学贡献方面是个好兆头。

The principal investigator and coworkers will study mathematical and computational issues of blind source separation algorithms of convolutive sound mixtures. The convolutive sound mixtures appear in any enclosed environment when there are multiple speakers. The goal is achieve dynamic separation of mixtures and source recovery adaptively with no prior knowledge of the environment, based on source independence and received data at multiple locations. Though the fast Fourier transform helps to localize the problem in the frequency domain, and separation is quite successful at each frequency with existing methods, permutation and scaling issues remain indeterminate and may greatly influence the separation quality. The investigators will use dynamically updated statistical signal information to fix permutation, and minimization of mixing filter lengths in the time domain to fix scaling. They will also study dynamic algorithms in the time domain by optimizing mixing filter lengths and source independence, as well as the dynamic stability and convergence of the recovered mixing filters by using probability theory.The project aims to develop algorithms to separate realistic sound mixtures, a task that machines (computers) are unable to perform as well as humans. The challenge, also known as the cocktail party problem, is fundamental to improving the quality of modern hearing devices. For example, hearing aids and cochlear implants are known to work well in quiet, however, they degrade rapidly when there are competing sound sources. Even for normal hearing people, it is difficult to carry out a conversation over a cell phone call that comes from a rather noisy location such as a restaurant. Understanding the mathematics of blind source separation and applying it to actual computation is a key step to solution. The project bodes well in generating broad impact and making mathematical contributions to the advancement of information technology and biotechnology.