Electromyography (EMG) signals carry information about the movements of skeleton muscles. EMG on-line processing and analysis can be applied to different types of human-machine interfaces and provide advantages to patient rehabilitation strategies in case of injuries or stroke. However, continuous monitoring and data collection produces large amounts of data and introduces a bottleneck for further processing by computing devices. Neuromorphic technology offers the possibility to process the data directly on the sensor side in real-time, and with very low power consumption. In this work we present the first steps toward the design of a neuromorphic event-based neural processing system that can be directly interfaced to surface EMG (sEMG) sensors for the on-line classification of the motor neuron output activities. We recorded the EMG signals related to two movements of open and closed hand gestures, converted them into asynchronous Address-Event Representation (AER) signals, provided them in input to a recurrent spiking neural network implemented on an ultra-low power neuromorphic chip, and analyzed the chip's response. We configured the recurrent network as a Liquid State Machine (LSM) as a means to classify the spatio-temporal data and evaluated the Separation Property (SP) of the liquid states for the two movements. We present experimental results which show how the activity of the silicon neurons can be encoded in state variables for which the average state distance is larger between two different gestures than it is between the same ones measured across different trials.