~0.5 % improvement

[jyrkialakuijala]

Score type	MSE	Min score	Max score	Mean score
Zimtohrli	0.077647743787295	0.608127552926780	0.829014091535645	0.728679787505096
ViSQOL	0.115330916105424	0.520833375452983	0.801480831107469	0.675101633981268
2f	0.129541391104905	0.484687555319526	0.797475783883375	0.661870345773127
PESQ	0.147425552045669	0.342342966279351	0.841271127756762	0.647128996775172
CDPAM	0.153471222942756	0.441558428344727	0.728779141125759	0.620699318941738
PARLAQ	0.185057687192323	0.445261140223642	0.784370761057963	0.587162756572532
AQUA	0.223207996944378	0.331645933512413	0.739286336419790	0.547804951221731
PEAQB	0.225217321572038	0.278744167467764	0.851011116004117	0.553935720513487
DPAM	0.315810440183130	0.186717781679534	0.690564701717118	0.460415212267967
WARP-Q	0.339686211572685	0.067600137543649	0.777119464646524	0.475793617709890
GVPMOS	0.412937133868407	0.006851162794410	0.783946603687895	0.412912222208318

[jyrkialakuijala]

It looks to me that the internal values are out of ranges where they are supposed to be and while this metric correlates fantastically with this corpora, there is likely a more fundamental issue here

it is possible that the loudness corrected spectra shouldn't be L2/NSIM matched but another distance metric should be used

eventual L2/NSIM matching likely drives the internal representations to be very strange and it is highly likely that the system generalizes relatively poorly in its current state to much louder or much more silent sounds as well as spectral extremes