Good ISTFT configuration for lower-resolution inputs

[janvainer]

Hi I am trying to train Vocos on features with lower time resolution.

The vanilla params are:

        n_fft: 1024
        hop_length: 256

what my experiments use:

        n_fft: 1920
        hop_length: 480

The result speaks but the voice quality is much worse compared to the vanilla training.
Could you please recommend some adjustments that might help with the output quality for the mentioned stft setup?