This study was designed to propose a feedback muscle force (or joint torque) control system for incompletely paralyzed patients whose muscle(s) was contracted by electrical stimulation and voluntary activation simultaneously. For this end, the joint torque needed to be monitored so as to adjust the electrical stimulation intensity accordingly. A new method named as the parallel filter algorithm was developed to separate the mixed electromyogram (MEMG) into the evoked electromyogram (EEMG) and voluntary electromyogram (VEMG). Contrary to the conventional filters, our algorithm consisted of (1) Fourier transformation, (2) multiplication of MEMG by two frequency functions, and (3) inverse Fourier transformation. The parallel filter algorithm enabled us to successfully extract EEMG and VEMG from MEMG, and the computation time was short enough to be applied to a real-time feedback control system. The radial basis function neural network algorithm was modified and employed to use EEMG and VEMG to estimate the joint torque components, evoked torque (ETorque) and voluntary torque (VTorque) respectively. These two torque components were then algebraically added to generate the total joint torque. This (estimated) joint torque was used for computing a new electrical stimulation intensity through a simple proportional-integral-differential controller. The experimental results suggested that the proposed parallel filter algorithm and the neural network algorithm could be embedded together with a controller into a real-time feedback joint torque control system.