A model of early auditory processing is proposed, in which each peripheral channel is processed by a delay-and-subtract cancellation filter. The delay is tuned automatically with a criterion of minimum power, independently for each channel. For a channel dominated by a pure tone, or a resolved partial of a complex tone, the optimal delay is its period. For a channel responding to harmonically-related partials, the optimal delay is their common fundamental period. Each channel is thus split into two subchannels, one filtered and the other not, either of which can be attended to depending on the task. Here, the model is used to explain the masking asymmetry between pure tones and noise. According to the model, a noise target masked by a tone is more easily detectable, thanks to cancellation, than a tone target masked by noise, as indeed is reported in the literature. The in-channel cancellation model is one of a wider class of models, monaural or binaural, that apply cancellation to suppress irrelevant stimulus dimensions so as to attain invariance to competing sources. Similar to occlusion in the visual domain, cancellation yields sensory evidence that is incomplete, thus requiring Bayesian inference of an internal model along the lines of Helmholtz's doctrine of unconscious inference.