The common problem of enhancing speech disturbed by background noise has been examined, under the assumption that no a priori knowledge about the noise is available. That is why we have built an algorithm trying to retain the part of the signal that follows a speech production model.