A framework for speech enhancement using extreme learning machines
Babafemi O. Odelowo, David V. Anderson · 2017
Neural networks have been widely applied in the enhancement of speech degraded by additive noise. In this paper, we study a mask-based speech enhancement framework employing the extreme learning machine (ELM). Experimental results show the proposed framework is extremely effective in suppressing additive noise when trained with small datasets. It is superior to a leading minimum mean-square error (MMSE) algorithm in both matched and mismatched noise conditions with just one hour of training data and performs consistently over a wide range of training conditions. In addition, the mask-based ELM approach is shown to be more effective than the spectral mapping approach.