Convolutional fusion network for monaural speech enhancement

Convolutional neural network (CNN) based methods, such as the convolutional encoder–decoder network, offer state-of-the-art results in monaural speech enhancement. In conventional large kernel size is often used to enhance model capacity, which, however, low parameter efficiency. This could be addressed by using group convolution, AlexNet, where convolutions are performed parallel each layer, before their outputs concatenated. However, with simple concatenation, inter-channel dependency information may lost. To address this, Shuffle re-arranges of concatenating them, taking part whole input sequence convolution. this work, we propose a new fusion (CFN) for enhancement improving performance, dependency, reuse and First, unit (GCFU) consisting standard depth-wise separable CNN reconstruct signal. Second, (full information) fed simultaneously two convolution networks parallel, re-arranged (shuffled) then concatenated, order exploit within network. Third, intra skip connection mechanism connect different layers inside encoder well decoder further improve performance. Extensive experiments show improved performance proposed method compared three recent baseline methods.