Abstract. This paper introduces gradient based method for robust assessment of the sparse pulse sources, such as motor unit innervation pulse trains in the filed of electromyography. The method employs multichannel recordings and is based on Convolution Kernel Compensation (CKC). In the first step, the unknown mixing channels (convolution kernels) are compensated, while in the second step the natural gradient algorithm is used to blindly optimize the estimated source pulse trains. The method was tested on the simulated mixtures with random mixing matrices, on synthetic surface electromyograms and on real surface electromyograms, recorded from the external anal sphincter muscle. The results prove the method is highly robust to noise and enables complete reconstruction of up to 10 concurrently active motor units.