Mandel Moshe ; Tal Or ; Adi Yossi

Abstract

We present AERO, a audio super-resolution model that processes speech andmusic signals in the spectral domain. AERO is based on an encoder-decoderarchitecture with U-Net like skip connections. We optimize the model using bothtime and frequency domain loss functions. Specifically, we consider a set ofreconstruction losses together with perceptual ones in the form of adversarialand feature discriminator loss functions. To better handle phase informationthe proposed method operates over the complex-valued spectrogram using twoseparate channels. Unlike prior work which mainly considers low and highfrequency concatenation for audio super-resolution, the proposed methoddirectly predicts the full frequency range. We demonstrate high performanceacross a wide range of sample rates considering both speech and music. AEROoutperforms the evaluated baselines considering Log-Spectral Distance, ViSQOL,and the subjective MUSHRA test. Audio samples and code are available athttps://pages.cs.huji.ac.il/adiyoss-lab/aero