Shrinkage based empirical mode decomposition for joint denoising and dereverberation

摘要

In this work, a novel algorithm has been proposed which can be employed for the noisy reverberant speech enhancement. Shrinkage based empirical mode decomposition will be employed using sub band processing. The method proposed here is a multistage algorithm using one microphone. Firstly, an EMD algorithm has been used to decompose a noisy reverberant speech signal into its oscillatory parts adaptively resulting in components named as Intrinsic Mode Functions (IMF). Then EMD based shrinkage method has been employed to the IMFs in order to reduce the noise, followed by the dereverberations of these denoised IMF components using Spectral subtraction. In the end we can achieve the enhanced signal via reconstruction mechanism from the processed IMFs. The main motivation behind this approach is the disproportional distribution of the noise and reverberations across the different IMF components. Therefore, we had used various levels of suppression in order to reduce the noise and reverberations across the different IMFs. On measurement basis of Signal to Noise ratio (SNR), the results were compared with a related state of art approach and an enhancement in the quality of speech signal was observed.