The aim of this paper is to present a novel method for simultaneous spike waveforms extraction and sorting from the raw recorded signal. The objective is twofold : on the one hand, to enhance spike sorting performance by extracting the spike waveforms of each spike and, on the other hand, to improve the analysis of the multi-scale relationships between spikes and LFP by offering an accurate separation of these two components constitutive of the raw micro recordings. Approach. The method, based on a Bayesian approach, is fully automated and provides a mean spike shape for each cluster, but also an estimate for each singular spike waveform, as well as the LFP signal cleaned of spiking activity. Main results. The performance of the algorithm is evaluated on simulated and real data, for which both the clustering and spike removal aspects are analyzed. Clustering performance significantly increases when compared to state-of-the-art methods, taking benefit from the separation of the spikes from the LFP handled by our model. Our method also performs better in removing the spikes from the LFP when compared to previously proposed methodologies, especially in the high frequency bands. The method is finally applied on real data (ClinicalTrials.gov Identifier : NCT02877576) and confirm the results obtained on benchmark signals. Significance. By separating more efficiently the spikes from the LFP background, our method allows both a better spike sorting and a more accurate estimate of the LFP, facilitating further analysis such as spike-LFP relationships.