An adaptive method for reliable and fast detection of muscle activity from surface electromyographic (sEMG) signals is introduced. The aim of this research was to minimize the delay of the onset and termination detection, while still retaining the reliability and simplicity of the detection algorithm. The proposed algorithm is based on a double-threshold detector. The algorithm applies the same principles as a constant false alarm rate (CFAR) processor that is often used to distinguish events from noisy environments with dynamic noise characteristics. The algorithm was tested with different noise conditions and frequencies. For each condition, a set of 1000 computer-simulated EMG signals were processed multiple times with different processing parameters in order to find the optimal settings for reliable muscle activity detection. The results for the detection delays were comparable to previously published results, and for low-noise conditions the detection worked without errors. The performance of the algorithm was verified using real sEMG signals. Performance in termination detection that has often been neglected in prior studies, is also reported. The results show that the method could be applied in the targeted real-time application: facial pacing.