A deep learning framework for audio restoration using Convolutional/Deconvolutional Deep Autoencoders

People communicate daily with their mobile phones and in some cases, the quality of the communication may be vital. Thus, there is a clear interest in improving the quality of communication in cases of low signal or interferences. This paper shows how deep learning techniques are used to restore audio files that simulate situations of background noise and loss of signal. Its main distinguishing feature is the direct use of the waveform instead of a spectrogram representation which lets the model be adapted to real-time communications or broadcasting. The results show that our proposal improves performance compared to Wave-U-Net. After restoring the audio, the difference between the original and the restored audio is, on average, less than 2%. In addition, a subjective test was carried out with 113 people who detected a significant improvement in the restored audio compared to the damaged one.