This thesis develops a unified computational pipeline to recover audio from degraded magnetic tapes using the Play It Again non-contact readout approach based on X-ray magnetic circular dichroism (XMCD). XMCD acquisitions produce sequences of high-dimensional detector frames in which the magnetization-dependent signal is weak and embedded in strong background and noise. The first contribution is an automated mask-estimation method that selects information-bearing pixels on the detector and enables a robust reduction from 2D dichroic frames to a 1D waveform. The mask is learned from a reference dichroic acquisition via unsupervised clustering on physics-inspired pixel features (intensity, local statistics, and gradient-based smoothness), with an optional translation (registration) step to compensate for small footprint drifts.