Iterative Thresholding-Based Spectral Subtraction Algorithm for Speech Enhancement

Abstract

Speech enhancement (SE) techniques rely much on noise estimation to remove noise from noisy speech. The simplest SE technique, spectral subtraction (SS), may perform much better than other SE techniques if the noise estimated is almost exact. In the proposed algorithm, SS is used because it can be easily implemented at a moderate computing platform like hearing aids or speech assistive devices. This paper presents iterative soft thresholding of the spectral domain for better noise estimation. The algorithm is based on the assumption that the speech signal is reasonably sparse in the spectral domain. The result shows that the proposed method performs better than IMCRA in terms of fwSNR and PESQ as quality measures, and in terms of STOI as an intelligibility measure under various noise environments like exhibition, babble, airport, and car at SNR level from −10 to 10 dB.