Electromyographic controls based on machine learning rely on the stability and repeatability of signals related to muscular activity. However, such algorithms are prone to several issues, making them non-viable in certain applications with low tolerances for delays and signal instability, such as exoskeleton control or teleimpedance. These issues can become dramatic whenever, e.g., muscular activity is present not only when the user is trying to move but also for mere gravity compensation, which generally becomes more prominent the more proximal a muscle is. A substantial part of this instability is attributed to electromyography's inherent heteroscedasticity. In this study, we introduce and characterize an adaptive filter for sEMG features in such applications, which automatically adjusts its own cutoff frequency to suit the current movement intention. The adaptive filter is tested offline and online on a regression-based joint torque predictor. Both the offline and the online test show that the adaptive filter leads to more accurate prediction in terms of root mean square error when compared to the unfiltered prediction and higher responsiveness of the signal in terms of lag when compared to the output of a conventional low-pass filter.