Radial basis function networks for noise reduction of speech

The paper compares eigen decomposition, linear and non linear radial basis function network based filter prediction, and signal embedding for the removal of additive white gaussian noise from speech. The embedding based predictor achieves improved noise reduction in comparison with eigen decomposition but does not perform as well as a non linear prediction. However, signal embedding offers a compromise between retaining the high frequency spectral information found in non linear prediction while achieving low frequency noise reduction comparable to eigen decomposition