EMD-based noise estimation and tracking (ENET) with application to speech enhancement

Speech enhancement from measured speech signals is fundamental in a wide range of instruments. It relies on a noise estimate which can be obtained using techniques such as the minimum statistics (MS) approach. In this paper, a novel approach for Empirical Mode Decomposition (EMD) based noise estimation and tracking (ENET) is presented with application to speech enhancement. Spectral analysis of non-stationary signals such as speech is performed effectively using EMD. The Improved Minima Controlled Recursive Averaging (IMCRA) that evolved from MS has been shown to be effective in non-stationary environments. ENET is able to use EMD in a novel way to estimate the noise spectrum more accurately than IMCRA and enhance speech more effectively than conventional log-MMSE approaches. A comparative performance study is included that demonstrates that it achieves improved speech quality than a conventional log-MMSE filtering approach with better noise estimation, even during periods of strong speech activity.