Comparison between soft and hard Thresholding on selected intrinsic mode selection

This paper uses an improved speech enhancement approach based on Empirical Mode Decomposition (EMD), Mode Selection approach and Thresholding technique through hard and soft functions. At first, by using a time decomposition called sifting process, the noisy speech signal is decomposed adaptively into intrinsic oscillatory components called Intrinsic Mode Functions (IMFs). Basically, the Modes Selection idea implies that the lower order IMFs (high-frequency modes) is mostly dominated by noise and the last ones (low-frequency modes) represent the most structures of the signal. Therefore, the denoised signal is partially reconstructed only by the low-frequency modes. Secluding the first IMFs may introduce a signal distortion rather than reducing noise. Our upgraded approach consists of thresholding the lower order modes using the soft or hard thresholding algorithm and that are added to the rest of IMFs. The simulations results show that the denoising with the hard function is more effective in removing the noise components then the soft one.