A denoising process models a noisy signal using classes and subclasses of symbol contexts. The process generates class count vectors having components that combine occurrence counts for different symbols in different contexts. Biases determined separately for each subclass and a fixed predictor indicate which symbol occurrence counts for different context are combined in the same component of a class count vector. For impulse noise, the bias for a subclass can be the average error that results when the fixed predictor predicts non-noisy symbols found in contexts of the context subclass. Denoising of impulse noise can select replacement symbols without matrix multiplication or a channel matrix inverse by evaluating distributions that result from subtracting error probabilities from probability vectors associated with respective contexts. Probability mass can be moved from adjacent components of the probability vector to assure that subtraction of the error probabilities leaves non-negative results.