Surface electromyography (sEMG)-based hand gesture recognition can enable intuitive prosthetic and human-machine control, but conventional time-domain features computed from the interference signal are only an indirect view of the underlying activation and may degrade for complex wrist motions. This thesis evaluates whether single-channel EMG deconvolution can provide reliable features for gesture classification while remaining practical for low-electrode systems. A publicly available multi-subject sEMG dataset (10 subjects, 64 channels, 6 gestures plus rest; 2 kHz) was used, but only 4 channels were selected to emulate wearable setups with few electrodes. Signals were segmented using 250 ms windows with 200 ms overlap (50 ms update), consistent with real-time controller delays.