Techniques for filtering noisy estimates to reduce estimation errors are described. A sequence of input values (e.g., for an initial channel impulse response estimate (CIRE)) is filtered with an infinite impulse response (IIR) filter having at least one coefficient to obtain a sequence of output values (e.g., for a filtered CIRE). The coefficient(s) are updated based on the sequence of input values with an adaptive filter, a bank of prediction filters, or a normalized variation technique. To update the coefficient(s) with the adaptive filter, a sequence of predicted values is derived based on the sequence of input values. Prediction errors between the sequence of predicted values and the sequence of input values are determined and filtered to obtain filtered prediction errors. The coefficient(s) of the IIR filter are then updated based on the prediction errors and the filtered prediction errors.