This work aims to investigate the potential of employing Denoising diffusion probabilistic models, commonly referred to as diffusion models, to revert a processed electric guitar recording to its original, unaltered form while retaining all the expressive elements of the performance such as dynamics and articulation. Specifically, a parallel dataset is constructed, containing both the unprocessed and processed versions of the guitar recordings, which is used for training a diffusion model. To preserve the expressiveness, the model is conditioned on the processed guitar recording when restoring the raw guitar sound. This research has the potential to enhance the accuracy of various music information retrieval tasks, such as automatic
music transcription.