The paper mentioned "hidden-dim activations are projected into a tensor h with n_fft + 2 channels"

I was wondering why you need n_fft + 2 number of channels (i.e. real numbers) to reconstruct the audio waveform. I understand that you would need only n_fft/2+1 number of fourier coefficients to represent the 'full' spectrum of an audio waveform of n_fft samples due to conjugate symmetry, but one or two of those coefficients will be real numbers depending on whether the number of audio samples is odd or even.

As a result, I believe you really theoretically only need n_fft number of channels/real numbers in fourier space to reconstruct the audio waveform, which makes sense since that is equal to the number of audio samples in the first place (i.e. no redundancy nor loss of information)