Acoustic Bandwidth Extension by Audio Deep Residual U-Net
Linlin Ou, Yuanping Chen · 2022
Mobile communication systems often rely on traditional channels with narrow bandwidth bottlenecks, reducing the audio quality. Due to the size of the network and the heterogeneity of devices, using high-quality audio codecs in real-world scenarios to transmit higher sample rate audio is difficult to practice. This paper proposes an approach by which communication nodes can extend the bandwidth of the band-limited input signal of a low sample rate codec, an audio synthesis neural network: ARUNet (Audio deep Residual U-Net), that integrates the advantage of residual learning, and U-Net is introduced to compensate for the high-frequency part of the live input audio clips. In addition, ARUNet has made outstanding achievements in various instrument solo data sets, with SNR close to 40dB and VGG distance close to 32. We explore the impact of our architectural design and demonstrate considerable improvements in terms of both perceptual and objective metrics.