Since the end of the 19th century, sound recordings have increased the possibilities and scope of musicology and music documentation. In Uruguay, the figure of Lauro Ayestarán stands out, who between 1943 and 1966 made more than 3000 field recordings throughout the national territory. These recordings are kept at the Centro Nacional de Documentación Musical (CDM) and suffer from a high level of noise that can make it difficult to appreciate, thus narrowing the access. Traditional methods for digital audio restoration are based on Digital Signal Processing techniques, such as Wiener filtering and autoregressive modelling. However, the significant progress brought by deep learning to computer vision and natural language processing has also extended to the audio domain, improving the state-of-the-art in problems like speech recognition and sound source separation. Consequently, some recent works have addressed audio restoration tasks using a deep learning approach, including bandwidth extension, and denoising. This work addresses the problem of noise reduction in analog tape recordings using a deep-learning approach. First, we build a data set of audio snippets of tape noise extracted from different functional tape equipment. Then, we train a deep-learning architecture. The data set of tape noise snippets and the trained deep-learning models are publicly available. In this way, we encourage the collective improvement of the data set and the broad application of the denoising approach by the audiovisual archives community. This interdisciplinary work is carried out by archive researchers from the CDM and signal processing researchers from the Universidad de la República (Uruguay).