We propose to implement speech enhancement by the regeneration of clean speech from a salient representation extracted from the noisy signal. The network that extracts salient features is trained using a set of weight-sharing clones of the extractor network. The clones receive mel-frequency spectra of different noi
تحميل البحث