Google I/O is a wrap! Catch up on TensorFlow sessions View sessions

Generate audio waveform from spectrogram using Griffin-Lim algorithm.

This is an adaptation of the method introduced in D. Griffin and Jae Lim, "Signal estimation from modified short-time Fourier transform,"

spectrogram A spectrogram with shape (None, nfft // 2 + 1). The first dimension depends on window and stride. The second contains half of the absolute values of the Fourier transforms because the FFT of a real signal is symmetrical.
nfft Size of FFT.
window Size of window.
stride Size of hops between windows.
iterations number of Griffin-Lim iterations. Change this to higher values if the reconstructed phase is not appropriate (i.e., the audio is coming out echoed or otherwise weird)

An 1-D waveform audio tensor.