Rapid Learning for Efficient Audio Acquisition in Constrained Environment

We propose an efficient audio acquisition scheme for constrained sensing environments based on Compressed Sensing (CS). Dictionary learning methods are designed to improve sparse representation, thereby enhancing the quality of the reconstructed signal from compressed sensing samples. Existing approaches generally involve a sparse coding stage that is computationally intensive, resulting in a time-consuming training process. In this work, we propose a fast dictionary learning scheme. First, a rapid and accurate voice activity detection module is employed to categorize the training data into two separate databases: one for voiced signals and one for unvoiced signals. Then, for each class, a labeled dictionary is learned using a greedy adaptive method that focuses only on the sparsity of the dictionary and avoids sparse coding techniques. Finally, a decision block is designed to determine whether the dictionary is voiced or unvoiced for the recovery stage of the compressive sensing procedure. Evaluation results show that the proposed method can achieve good performance in terms of several objective measures, including perceptual evaluation of speech quality (PESQ), short-time objective intelligibility (STOI), and segmental signal-to-noise ratio (SSNR), especially in noisy conditions. Overall findings highlight the merit of our presented system in terms of signal quality, learning speed, and efficient dictionary storage.