Online sound restoration system for digital library applications

Abstract

Audio signal processing algorithms were introduced to the new online non-commercial service for audio restoration intended to enhance the content of digitized audio repositories. Missing or distorted audio samples are predicted using neural networks and a specific implementation of the Jannsen interpolation method based on the autoregressive model (AR) combined with the iterative restoring of missing signal samples. Since the distortion prediction and compensations algorithms are computationally complex, an implementation which uses parallel computing has been proposed. Many archival recordings are at the same time clipped and affected by wideband noise. To restore those recordings, the algorithm based on the concatenation of signal clipping reduction and spectral expansion was proposed. The clipping reduction algorithm uses an intelligent interpolation to replace distorted samples with the predicted ones based on learning algorithms. Next, spectral expansion is performed in order to reduce the overall level of noise. The online service has been extended with some copyright protection mechanisms. Immunity of watermarks to the sound restoration is discussed with regards to low-level music feature vectors embedded as watermarks. Then, algorithmic issues pertaining watermarking techniques are briefly recalled. The architecture of the designed system together with the employed workflow for embedding and extracting the watermark are described. The implementation phase is presented and the experimental results are reported.