Abstract: Speech and audio signal processing frequently requires to recover a time-domain signal from the magnitude of a spectrogram. Conventional methods inversely transform the magnitude spectrogram ...