The compound surface electromyograms (EMGs) recorded from patients with dystonia commonly contains superimposed bursting and tonic activity representing various motor symptoms. It is desirable to differentially extract them from the compound EMGs so that different symptoms can be more specifically investigated and different mechanisms revealed. A non-linear denoising approach based on wavelet transformation was investigated by applying soft thresholding to the wavelet coefficients. Thresholds were determined according to three different principles and two models. Different techniques for wavelet shrinkage were investigated for separating burst and tonic activity in the compound EMGs. The combination of Stein's unbiased risk estimate principle with a non-white noise model proved optimal for separating burst and tonic activity. These turned out to be exponentially related; and the temporal relationships between antagonist muscle contractions could now be seen clearly. We conclude that adaptive soft-thresholding wavelet shrinkage provides effective separation of burst and tonic activity in the compound EMG in dystonia. This separation should improve our understanding of the pathophysiology of dystonia.