Blind adaptive filtering of speech from noise of unknown spectrum using a virtual feedback configuration

The paper describes a blind adaptive filter (BAF) of speech from noise where neither speech nor noise are accessible, nor are their parameters known. The BAF employs prior knowledge on the nonstationarity structure of human speech. The noise is assumed to be quasi-stationary over a longer interval than any speech phoneme. The system employs an identifier that is followed by a speech/noise parameter-separator based on stationarity features. It then employs a virtual feedback loop where a cost-functional is minimized to jointly minimize the stationary part of the output while maximizing its nonstationary part. SNR improvements for various noises range from 14 to 36 dB for -20 dB SNR inputs