the griffin lim algorithm works to reconstruct audio from a magnitude spectrogram, but i don't understand how reusing the same spectrogram every iteration with noise can produce decent sound
How from the random phase is ordered to be heard well?
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
Sound is just the combination of multiple sin waves at different frequencies and phases. These phases determine how those sin waves combine together, i.e. if frequencies are resonant or eliminate each other. A shift in phase could change the magnitude of certain frequencies. Taking advantage of this property, the Griffin-Lim algorithm tries to iteratively optimize the spectrogram phases so that the reconstructed magnitude is as close as possible to the original magnitude. Hope this could answer your question.
