Automatic urban sound classification is a desirable capability for urban monitoring systems, allowing real-time monitoring of urban environments and recognition of events. Current embedded systems provide enough computational power to perform real-time urban audio recognition. Using such devices for the edge computation when acting as nodes of Wireless Sensor Networks (WSN) drastically alleviates the required bandwidth consumption. In this paper, we evaluate classical Machine Learning (ML) techniques for urban sound classification on embedded devices with respect to accuracy and execution time. This evaluation provides a real estimation of what can be expected when performing urban sound classification on such constrained devices. In addition, a cascade approach is also proposed to combine ML techniques by exploiting embedded characteristics such as pipeline or multi-thread execution present in current embedded devices. The accuracy of this approach is similar to the traditional solutions, but provides in addition more flexibility to prioritize accuracy or timing.