A Phase Aware Audio Declipping Method Using BandSplit Recurrent Neural Network in Heavily Noisy Environments

In digital audio systems, clipping occurs when the amplitude of a signal exceeds a threshold, leading tosignal distortion and unpleasant noise to the listener. Therefore, a declipping process is required to recover theclipped portion and reconstruct the signal. In conventional deep neural network-based audio enhancement methods, the focus has primarily been on restoring the magnitude spectrum, but recent studies indicate that enhancing the phase spectrum is also crucial for improving quality. In this paper, we propose an audiodeclipping method based on the BSRNN(band-split recurrent neural network) that utilizes phase-based features such as instantaneous frequency deviation (IFD), or applies the neural vocoder HiFi-GAN (generativeadversarial network for efficient and high-fidelity speech synthesis) in the post-processing stage to improve theobjective quality of signal. The experimental results show that the proposed method outperforms theconventional magnitude spectrum-based enhancement method and the DCCRN model-based declipping methodaccording to DNSMOS P.835 OVRL score.