Speech enhancement employing loudness subtraction and oversubtraction

Speed enhancement approaches based on loudness substraction and over-substraction is presented. The spectral over-substraction method is proposed to provide further improvement and implements a SNR dependent substraction factor that applies a higher substraction factor in the low SNR frames and vice versa. The relative quantity of a speech signal depends on the difference in the loudness domain between the signal and the reference speech signal in the PESQ measure. The proposed loudness over substraction model substracts a portion of the average loudness of the noise from the noisy speech signal that always show fluctuations around average that may lead to large noise residues in the enhanced signals. The approaches in the advanced domain results in improved Segmental SNR, improved PESQ scores, and less distortion compared to the corresponding algorithms in the spectral domain.