Adaptive Speech Enhancement and Signal Separation Using Robust Neural Network Estimation Techniques

点击翻译按钮获取中文摘要

The objective of this project is the development of robust methods for enhancing speech in real-world noisy environments. A neural network approach is developed for time-domain enhancement. The availability of only the noisy signal is assumed, in contrast to previous neural network approaches. Effectively, a sequence of neural networks is trained on the specific noisy speech signal, resulting in a nonstationary model that can be used to remove noise from the given signal. Two related frameworks are developed: (1) a neural state-space approach based on nonlinear extensions to Kalman filter theory and (2) a direct neural mapping from noisy speech to enhanced speech. This research represents fundamental enabling technology for interactive systems requiring robust speech communications. More traditional methods are often effective under ideal situations, but tend to degrade for real world nonstationary noise sources.