Multiresolution STFT audio processing usually has the problem how to detect and separate transients from steady-state signals. We propose a method to avoid this issue by initializing the phase estimation of the longwindow STFT with the result of the short-window STFT and vice versa. The reason behind this approach is that the better temporal resolution of the short-window STFT moves information about the temporal behavior of the signal from the phase spectrum to the magnitude spectrogram, making it accessible to the phase estimator in the initialization step. An evaluation shows the advantage of this method compared to the previously used approaches.
This paper costs $33 for non-members and is free for AES members and E-Library subscribers.