A Fully Convolutional Neural Network For Complex Spectrogram Processing In Speech Enhancement

In this paper we propose a fully convolutional neural network (CNN) for complex spectrogram processing in speech enhancement. The proposed CNN consists of one-dimensional (1-d) convolution and frequency-dilated 2-d convolution, and incorporates a residual learning and skip-connection structure. Compared with the state-of-the-art, the proposed CNN achieves a better performance with fewer parameters. Experiments have shown that the complex spectrogram processing is effective in terms of phase estimation, which benefits the reconstruction of clean speech especially in the female speech case. It is also demonstrated that the model yields a convincing performance with small memory footprint when the number of parameters is limited.