Speech enhancement using orthogonal matching pursuit algorithm

This paper proposes a novel approach for speech enhancement based on compressed sensing (CS) theory. Each frame of noisy speech signal is sparsified firstly by using discrete cosine transform (DCT). Then we divide each frame into the noisy sub-frame and the clear sub-frame with a soft thresholding method to obtain the threholded DCT coefficients of the noisy sub-frames. After that, the partial Hadamard ensemble is used as a sensing matrix to achieve compressive measurement of the DCT coefficients of noisy sub-frame. Finally, We use the orthogonal matching pursuit in order to recover the de-noised speech signal from noisy sub-frame. Both objective and subjective experiments are employed to compare the proposed approach with the subspace method and the spectral subtraction method. Experimental results shows that proposed method outperforms other methods with the highest PESQ, ABX and MOS score for Gaussian white noise and most kinds of colour noise.