This paper deals with a technique to remove noise from a voicelike signal. It is assumed that the listening device has attained a certain familiarity with the source of the signal. In fact, the first part of the paper deals with certain measurements to be performed to identify and measure some of the characteristics of the source. Once these characteristics are determined, the source produces a sequence of signals (most of which were not used during the familiarization process) that are mixed with noise. The task now is to remove the noise and identify the signals. A model of voice production (described at an earlier meeting of the ASA) has been used, which begins with a nonlinear section followed by a linear section. The nonlinear section characterizes the source and the linear section determines various vowellike sounds. It has been assumed in this paper that the nonlinear section does not change (which would hold the pitch constant) and that the linear section is characterized by two unknown formants. Within the limitations imposed by this rather simplistic model, a large amount of noise can be removed from the signal.