Developing tools for the recovery of historical audio recordings

Link

Magnetic tape recordings hold historically irreplaceable audio, yet many recordings are unplayable due to physical and chemical degradation, which pose challenges to their preservation. This thesis develops tools for recovering audio from data obtained using non-contact readout techniques, such as X-ray Magnetic Circular Dichroism (XMCD) and Tunneling Magnetoresistance (TMR) sensors.

Signal denoising was tackled using three methods: Extended and Unscented Kalman Filters for model-based denoising, with the UKF producing cleaner spectrograms than the EKF, but both requiring extensive tuning; Spectrogram soft-masking with Griffin-Lim phase reconstruction, which effectively removed broadband noise for isolated tones but introduced artifacts for complex sweeps; and a two-stage U-Net (Moliner et al.) pretrained on gramophone recordings, which revealed obscured harmonic content in a Beethoven excerpt, illustrating the promise of deep learning for this task, but still needs further validation and fine-tuning.

A recording-to-readout simulation pipeline was implemented. It incorporates the Karlqvist field, AC bias, hysteresis behavior via Jiles-Atherton and Preisach models, includes Mallinson-type noise, and represents the XMCD beam as a skewed Gaussian convolution. Experimental hysteresis loops were fitted using both models, with the Preisach providing superior accuracy (NRMSE 3.83%) compared to JilesAtherton. This framework enables the study of particle statistics and measurement parameters on the recovered signal. These tools form the basis for future reconstruction of real historical audio from degraded tapes using XMCD measurements. The simulation tools were validated on experimental XMCD scans of 1 kHz, 1-20 kHz and musical excerpts, confirming the feasibility of non-contact audio recovery