Griffin-Lim phase reconstruction using short-time Fourier transform with zero-padded frame analysis
Yukoh Wakabayashi, Nobutaka Ono · 2019
In this paper, we present the short-time Fourier transform (STFT) with zero-padded frame analysis to introduce frequency redundancy into a time-frequency representation, and we investigate its application to phase reconstruction by the Griffin-Lim algorithm. Recent studies on phase reconstruction have suggested that the use of a small STFT frame shift improves the performance of phase reconstruction techniques, which implies that increasing the temporal redundancy of the time-frequency representation makes phase reconstruction easier. Motivated by this, we consider the STFT with zero padding to increase the redundancy of the STFT along the frequency axis. The linearity of this transform and its inverse enables the use of the Griffin-Lim algorithm on the time-frequency domain. We evaluate the performance of phase reconstruction using the STFT with and without zero padding using the spectral distance and perceptual evaluation of speech quality as the criteria. The experimental results show that increasing the frequency redundancy with zero padding improves the phase reconstruction performance similarly to using a small frame shift.