When analyzing brain activity such as local field potentials (LFP), it is often desired to represent neural events by stereotypic waveforms. Due to the non-deterministic nature of the neural responses, an adequate waveform estimate typically requires to record multiple repetitions of the neural events. It is common practice to segment the recorded signal into event-related epochs and calculate their average. This approach suffers from two major drawbacks: (i) epoching can be problematic, especially in the case of overlapping neural events and (ii) variability of the neural events across epochs (such as varying onset latencies) is not accounted for, which may lead to a distorted average. In this paper, we propose a novel method called adaptive waveform learning (AWL). It is designed to learn multi-component representations of neural events while explicitly capturing and compensating for waveform variability, such as changing latencies or more general shape variations. Thanks to its generality, it can be applied to both epoched (i.e., segmented) and continuous (i.e., non-epoched) signals by making the corresponding specializations to the algorithm. We evaluate AWL's performance and robustness to noise on simulated data and demonstrate its empirical utility on an electrophysiological recording containing intracranial epileptiform discharges (epileptic spikes).