Dantzig selector for audio data reconstruction

Abstract

The paper aims at applying a sparse representation technique named Dantzig selector (DS) for audio signal decoding and reconstruction. Dantzig selector is compared with another sparse signal representation variant called gradient projection for sparse reconstruction (GPSR). The comparison is performed in terms of their ability to efficiently compact different types of signals when those signals are represented thorough specific numbers of measurements, employing mean squared error as similarity measure between the original and reconstructed signal. For synthetic spikes data and moderate number of measurements, DS leaded to more accurate signal reconstruction, while the GPSR became more efficient with an increase of the measurements number. However, for the audio data, DS constantly outperformed GPSR for varying number of measurements. Moreover their efficiency is assessed with respect to the number of non-zero values (cardinality). As the signal becomes denser, DS shows superior performance over GPSR for synthetic spikes data.