Exploring Audio Compression as Image Completion in Time-Frequency Domain

Audio compression is usually achieved with algorithms that exploit spectral properties of the given signal such as frequency or temporal masking. In this paper we propose to tackle such a problem from a different point of view, considering the time-frequency domain of an audio signal as an intensity map to be reconstructed via a data-driven approach. The compression stage removes some selected input values from the time-frequency representation of the original signal. Then, decompression works by reconstructing the missing samples as an image completion task. Our method is divided into two main parts: first, we analyse the feasibility of a data-driven audio reconstruction with missing samples in its time-frequency representation. To do so, we exploit an existing CNN model designed for depth completion, involving a sequence of sparse convolutions to deal with absent values. Second, we propose a method to select the values to be removed at compression stage, maximizing the perceived audio quality of the decompressed signal. In the experimental section we validate the proposed technique on some standard audio datasets and provide an extensive study on the quality of the reconstructed signal under different conditions.