Speech Super Resolution and Noise Suppression System Using a Two-Stage Neural Network

Super-resolution (SR) and noise suppression have been a hot research topic in the field of speech processing. Speech super-resolution, also named bandwidth extension (BWE), aims to extend the bandwidth of narrowband speech and improve its clarity, and noise suppression focuses on removing background noise from speech. Most of the past researches have performed these two tasks separately, while in the real world, bandwidth loss and noise exist almost simultaneously. Therefore, it is important to treat super-resolution and noise suppression jointly. In recent years, deep learning has made a big splash in the field of speech processing, and we see the promise of using deep learning techniques to solve this problem. In this paper, we propose a joint speech super-resolution and noise suppression method based on deep learning methods. Specifically, we use two networks to handle these two single tasks separately and then cascade them, we use multiple loss functions as well as multiplexing of the spectrogram, and the experiments show that our method outperforms the baselines.