Investigating the effect of phase reconstruction on the bandwidth extention on audio signal using DNNs

In this paper, the effect of phase reconstruction for bandwidth extension has been investigated using two different methods. In most signal processing techniques, given that the human hearing system is not susceptible to high phase degradation, the phase is abandoned and also extracted from the test signal at the test phase. In bandwidth extension, there is no direct use of the phase of the input signal for high-band because the phase of the input is only the phase associated with the low-band and the high-band phase must be estimated. In the present paper, the effect of phase reconstruction on two different methods has been investigated, one of them is the use of deep neural networks (DNNs) and the other is the use of a phase reconstruction algorithm. The results show that phase estimation using deep neural networks is not computationally efficient due to the increase in the input dimension of the DNNs.