Method and means for adaptively filtering near-stationary noise from an

Abstract

An input signal containing information such as speech or music as well as near-stationary noise is applied in parallel to a noise-analysis circuit and a noise-reduction circuit, each of which comprises a plurality of bandpass filters covering the range of frequencies associated with the information. The absolute value, or a function thereof, of the output of each bandpass filter in the noise-analysis circuit is produced and smoothed. The presence of near-stationary noise in the input signal is determined by examining the nature of the smoothed signal in each band assuming noise has a frequency spectrum which does not vary with time or varies only within a narrow range over a predetermined period of time with respect to the spectral parameters of the information signal. If noise is detected, the noise-analysis circuit identifies spectral parameters of the information and/or noise in each band using the smoothed signal therein. In the preferred embodiment of the invention, the bandpass filters of the noise-reduction circuit have gain elements that are adjusted in accordance with the identified parameters to minimize, under some continuous minimization criterion, the effect of the noise in the input signal thus enhancing intelligibility of the information therein. Minimization can be such that the gain-to-parameter relationships are similar to those in Weiner or Kalman filtering theory with a-priori knowledge of the noise, or of the noise and information, except that in this case, a-priori knowledge of the noise is acquired via identification and is not preassumed.