Wiener filtering is used to reduce the quantization error of the DCT transform coefficients. The process is applied in the transform domain with a sub-band decomposition structure. That is, the corresponding coefficients from each block are grouped together. To model the spectrum of the transform coefficients, two approaches are taken. The first one is to directly estimate the spectrum of the whole sequence. The second one is to remove the dependence among the sequence and model the mean-removed sequence as white noise. With the estimated spectrum, Wiener filter is then used to extract the original coefficients from the quantized ones.