The determination of propulsive and recovery phases of manual wheelchair propulsion are important in estimating the efficacy of manual propulsion. With the actual potential of collection of database by different research team, manual determination of propulsive and recovery phases of each wheelchair cycle represents a tremendous amount of time. Automatic detection with only a single threshold and low-pass filtering is insufficient to detect initial and terminal phases during propulsion. The purpose of this work is to present a new adaptive method based on short-time spectral analysis (STSA) of forces and moment acting at the handrim level. Eight able-bodied subjects were asked to propel on ergometer with 3 slopes, while forces and moments were recorded with an instrumented rear wheels. The STSA was applied to raw signal and automatic detection was based on priori and a posterior probability of absence and presence of propulsive signal. Better results are obtained with the new adaptive method when compared to the classical method with one single threshold and one cut-off frequency per signal. This method will prove useful when monitoring manual propulsion for a long term, and with different types of propulsion and classification of pattern of daily activities.