Experimental study on deep learning for spectrum reconstruction of damaged audio signals

This thesis presents an experimental study on the application of deep learning techniques for the spectrum reconstruction of damaged audio signals. Magnetic tapes used in 8mm film recordings often suffer from distortions where the audio signal becomes muffled in certain moments, while background noise remains constant. This degradation results in the loss of higher frequencies and the overall clarity of the audio, making the reconstruction of the original signal challenging. The goal of this research is to develop a method using deep learning models to predict and restore the missing or degraded portions of the audio signal. By analyzing the intact parts of the recording, the model learns to estimate the lost frequencies and reconstruct the signal, aiming for a continuous and coherent sound restoration.