In this paper we present an efficient method to generate realistic simulations of extracellular recordings. The method uses a hybrid and computationally simple approach, where the features of the background noise arise naturally from its biophysical process of generation. The generated data resemble the characteristics of real recordings, as quantified by the amplitude and frequency distributions. Moreover, we reproduce real features that are specially challenging for the analysis of extracellular data, such as the presence of sparse firing neurons and multi-unit activity. We compare our simulations with real recordings from the human medial temporal lobe and exemplify their use for testing spike detection and sorting algorithms. These results show that this technique provides an optimal scenario for generating realistic simulations of extracellular recordings.