An entropy-based neural fuzzy network estimation for speech enhancement

This paper discusses the problem of speech enhancement. The noise level varies in the procedure of recording due to speed change and moving environment. This condition usually results in wrong noise estimation and wrong speech enhancement process. To overcome these problems, one entropy based parameter (MS-entropy) and two energy-based temporal variation parameters (SV-MiFre & LV-MiFre) are proposed to improve the noise level estimation in the speech segment. Since the entropy based parameter can calculate the uncertainty of spectral magnitude, and the energy based temporal variation parameters can process the spectrum energy, the noise level estimated by the proposed method is more precise than that estimated by the pure spectrum energy method by C.T. Lin (2003). In addition, we use the self-organizing neural fuzzy network to avoid the need of empirically determining these noise estimation rules.