The magnitude error (DFmg) does not account the “magnitude differences” you mention (caused by the phase errors) because magnitude components of the output signal are not amplified due to the phase errors. For example the magnitude diffrograms of #9038 do not show inaccuracy of low frequencies...
I don't know exactly how various phase shifts are perceived by the ear and I also - as many - think that the low-freq phase shift is not too important. But I also know that if we replace all the phases of some music signal with the random ones, we will get just a noise shaped according to a freq...
Yes, I attenuated the null to the same perceived loudness in order to evaluate the type of the degradation. The M0's null is attenuated much more (you can listen the noise).
Here are the results of testing #9038 and M0pro with the multitone signal:
DFwf = -34.12 dB
DFmg = -52.27 dB
DFph = -45.10 dB
DFwf = -75.68 dB
DFmg = -75.57 dB
DFph = -63.91 dB
... and with the classical piece from the GearSpace forum:
DFwf = -36.21 dB
DFmg = -55.70 dB...
The test vectors used for building df-slides of DAPs:
t-signals + m-signal(Diversity) @44.1/16 - http://soundexpert.org/vault/media/combo-vector-7.2%2844-16-S%29-play.flac
t-signals @96/24 - http://soundexpert.org/vault/media/t-vector-7.2%2896-24-S%29-play.flac
I didn't write a paper about the algo but the Matlab code is open (http://soundexpert.org/articles/-/blogs/visualization-of-distortion#part3) and you can always check that algo by feeding exactly the same signals as input and output into diffrogram utility. In order to have 0.1dB accuracy of Df...
The null-difference has one disadvantage - it depends on the level of a reference signal. Df is a ratio of the null RMS to the level of reference signal. So, Df is a relative parameter and has exactly the same physical meaning as simple RMS of null difference.
Yes, that is the difference between df-metric and all others )) - it does not account psychoacoustics, instead it defines a threshold of transparency easily measurable and achievable at the current level of chip manufacturing. At that level the psychoacoustic properties of a degradation are not...
The “efficient resampling method” was developed. The time warping algo does it with any predefined accuracy. For example White Noise can be shrinked/stretched with df=-100dB accuracy in the current version. So, the origin of the time inconsistency in #9038 is not the time warping algo.