Accurate onset detection of the voluntary response is a prerequisite in reaction time studies when used in investigations on human motor control. The detection algorithm required performs a transformation of the continuous physical signal (e.g., force, movement) containing the response into a discrete event from which the reaction time (RT) is derived. Therefore, RT always comprises both the cognitive and/or motor delay component imaging the duration of the initiation process conducted by the sensorimotor system and, in addition, some spurious delay caused by the detection algorithm. As a standard method, onset detection is realized by the measurement device itself (e.g., the release of a micro switch) by defining the response onset as the point where the observed signal passes a certain threshold. Thus, weak and abnormal response profiles which are typical for a variety of central motor disorders (e.g., Parkinson's disease) may introduce high RT variability as well as systematic errors. The aim of this study was to improve accuracy of onset detection by application of an appropriate filter to the measured signal before entering the final decision stage. Three algorithms (lowpass differentiator, inverse filter, linear autoregressive (AR) predictor) were implemented and tested on simulated and real data under both on-line and off-line conditions with special interest to the influence of quasi-periodic background activity like tremor. It is shown that a significant improvement in onset detection accuracy, compared with the simple switch, can be achieved by using appropriate low order adaptive filters with the AR-predictor being the most efficient solution.