Multi-channel audio statistical restoration

Phonograph record is an analog sound storage medium that has played an important part in sound history. Modulated spiral groves on the disc are usually inevitably damaged by scratches and dust, which lead to noises with different statistical characteristics. Among those different type of noises, there is one common click-shape noise that have some degree of correlation between channels. Its origins have been explained as a deviation of the probe needle means it shifts closer to one wall but further from another with similar distances. Our aim is to restore this type of noises on dual mono audio signals with a statistical method. There are previous researches on restoring single-channel noisy data based on generative models. However multi-channel noisy data contain more audio information than single-channel data, due to its channel correlation. To exploit this property, we propose multivariate Gaussian models for both signal and noise models. Then we derive the Maximum A Posteriori (MAP) estimation of the underlying data. Additionally, we also use the Maximum Likelihood method to optimize model parameters. In the end we compare the restoration performances between our model and the baseline model on both synthetic and real-life noisy music data.