Speech Signal Recovery Using Block Sparse Bayesian Learning

Compressed sensing is based on the recovery of original signal from the low-quality and incomplete samples. Recently, ℓ _1 -norm is used for the estimation of signal elements from the underdetermined set of equations. In this paper, we propose a technique for speech signal recovery called block sparse Bayesian learning. The proposed technique is applied over the random set of speech samples and acquired better performance as compared to ℓ _1 -based recovery. Apart from the proposed recovery technique, this work is also intended to develop a trained and efficient sampling matrix through offline training. In this work, we apply structural similarity index as a metric to compare the performance of the proposed technique with an existing ℓ _1 based recovery. Sparse Bayesian learning and ℓ _1 -norm recovery are applied over the selected audio files from the datasets. The dataset consists of speech signals from three different languages: Urdu, Pashto and English. Structural similarity between the recovered and original speech signals is used as a metric to compare the performance of BSBL with ℓ _1 -norm minimization. The comparison based on structural similarity index shows the effectiveness of the proposed technique.