A Lightweight Efficient U-Net Model for Audio Super-Resolution

Audio super-resolution involves enhancing the quality and resolution of audio signals, particularly by improving their frequency content and overall clarity. In this paper, we explored the design and evaluation of a lightweight deep learning model based on the popular U-Net architecture for the audio super-resolution task. The proposed model consists of two identical parallel subnetworks: one for reconstructing the full spectrogram of the super-resolution signal and another for reconstructing its high-frequency content. The subnetworks are simple stacks of encoder and decoder blocks with skip connections. Our evaluation demonstrated that the proposed model can achieve comparable or better performance in terms of the objective quality of the upscaled audio signal when compared to other models while being 13 times smaller.