Myo-electric signals have been widely used in human-machine interfaces because these biosignal directly reflect human intentions to robots. The major difficulty of applying these biosignal in a pattern recognition system in real time is that they are unstable and vary in time. This instability occurs outside of the steady state of the signal, at the beginning and the ending of the motions. For real-time application users, the errors at the beginning of motion can lower the credibility of a pattern recognition system. In this sense, precise classification is the most significant factor for the system; thus the classification accuracy has higher priority compared to other factors. Generally, a trade-off relationship between the time delay of control commands and the classification accuracy has been known for sEMG users. Since parameters for signal processing can alter the sensitivity(time delay and accuracy) of the system, this study investigates limitations of a pattern recognition system due to transient-state errors. In particular, the performance of the system is analysed with respect to the analysis window size and the voting size of classification. Through an off-line simulation, we propose useful guidelines for the analysis window size and voting size in myoelectric signals for real-time applications.