Signal Separation and Enhancement using Multichannel/Multimodal Side Information

点击翻译按钮获取中文摘要

The aim of signal separation is to recover source signals from one or more signal mixtures. These mixtures can come from a multitude of domains, including speech mixtures for the so-called cocktail party problem, musical recordings of multiple instruments and/or vocals, electroencephalograph (EEG) signals from the electrical activity of the brain, maternal/fetal electrocardiograph (ECG) mixtures when measuring surface potentials for fetal monitoring, electromyography (EMG) signals from mixtures of motor unit action potentials in muscles, various astronomical datasets with mixtures of weak astronomical sources versus background, and radar or sonar systems. The focus for this proposal is to research and develop advanced source separation techniques using multichannel and multimodal metadata as side information to improve the quality in source separation. The signal processing research will concentrate on incorporating multichannel and side information into the source separation problem and in evaluating the impact of such side metadata information. Of particular interest is to extend non-negative matrix/tensor factorization techniques and incorporate side-information and convergence constraints through machine learning and deep learning, initially via convolutional neural networks. These signal separation characteristics will be research and studied under various applications, including speech, vocals/instrument separation, and EMG signal decomposition, but, the key research focus will be on the core components of source separation in multichannel mixtures with side metadata information. Some of the anticipated outcomes are as follows. First, in the vocal/instrument separation space, we expect to be able to take a musical recording, likely as a stereo multichannel recording, along with side metadata, which could include the musical score, sound samples of the instruments in the performance, an example performance of the piece by a secondary musician, a list of lyrics, a MIDI file, or other types of metadata, and be able to separate the individual instruments in the recording to either isolate or suppress individual instruments. In the EMG signal decomposition space, we expect to be able to take multichannel EMG recordings and determine firing times of individual motor units in the muscle and the corresponding shape of the motor unit action potentials. In the fetal ECG separation space, we expect to be able take a multichannel mixtures containing the stronger maternal ECG and weaker fetal ECG, and be able to suppress the stronger maternal ECG signal in an improved fashion. The proposed core research in signal separation will have wider applicability, where these examples would be used to validate the techniques and determine other limitations that may present themselves in one domain but not another.