question about loudness based loss

[faroit]

RAVE/rave/core.py

Line 130 in 54c6106

x = torch.log(x + 1e-7) + self.a_weight

I understand that the loudness based distance is just an approximation but I wonder why its added to the log spectrogram instead of mulitplying. Furthermore, what about using a time-domain prefilter as implemented by @csteinmetz1:

https://github.com/csteinmetz1/auraloss/blob/e732234398ada867138be634dbf66f40360461a2/auraloss/perceptual.py#L124-L129

and then implement it like:

mse(self.log_stft(time_domain_prefilter(x)), self.log_stft(time_domain_prefilter(y)))

can you comment on some pro and cons of either way?

[domkirke]

in this code a_weight are also in db, such that the multiplications becomes an addition.

[faroit]

@domkirke ahhh... missed that 🤦 thanks