It is of a major interest for contemporary and emerging wireless communication systems to be able to compute error rates fastly, efficiently and in real time. We propose a novel kernel-based Symbol Error Probability (SEP) estimation approach for Quadrature Amplitude Modulation (QAM) systems. The received soft observations are partitioned using an unsupervised stochastic expectation maximisation algorithm. The SEP is estimated using kernel-based probability density function estimate of modified versions of the received observations. Simu- lation results involving 4-QAM transmissions over a frequency- selective channel showed, for practically equivalent estimation accuracy, drastical computational cost reductions compared to the conventional way of estimating the SEP.