Blind adaptive filtering of unknown speech from unknown noise in a single receiver situation

Describes a blind adaptive filtering approach to filtering speech from noise when no access is available to speech or to noise alone and no parameter or spectral information is available on either. The approach is concerned with a single source. The solution to this problem involves a priori assumptions on features such as relative degree or stationarity or model order rather than on a parameter or spectrum. The feature information is incorporated in the received signal identification scheme. The resultant thus separated parameters are then employed to tune a filter which is based on a modified Kalman or Wiener filter whose output is passed through a network that biases that output according to the distribution of information in speech over the speech spectrum. It has been found that without the latter nonlinear feature speech intelligibility is hardly enhanced even when the SNR is improved. Real time results involving intelligibility tests on single syllable word lists show dramatic improvement in intelligibility.<
>