... but the result (according to df-metric) is very similar with this m-signal. Can you tell them apart by listening?
I think so, but I know that won't suffice so I guess it's ABX test time. :-( foobar is playing up, so I'll try the "ABX Comparator" facility in DeltaWave.
Hmm, first few trials didn't go too well but... selecting "X" (even with "Restart on switch" deselected) seems to result in the file starting from the beginning again (hmm... why?) placing too much dependency on short-term memory. OK, so removing the first 1.6 seconds of the file and starting again...
Trial 1, user: B actual: B 1/1
Trial 2, user: A actual: B 1/2
Trial 3, user: A actual: A 2/3
Trial 4, user: A actual: A 3/4
Trial 5, user: A actual: A 4/5
Trial 6, user: B actual: B 5/6
Probability of guessing: 10.9%
A=Reference,B=Comparison
Test type: ABX
Where:
Reference: Softube Trident A-Range - Remove First 1.6s.flac[?] 3942540 samples 44100Hz 24bits, ch=0, MD5=00
Comparison: Soft Clip Test - Remove First 1.6s.flac[?] 3942540 samples 44100Hz 24bits, ch=0, MD5=00
ABX tests are no fun, and I'm only using my "desktop" speakers. (Bose Computer MusicMonitor.) I can't be bothered to do any more trials right now.
I would also prefer if the software allowed one to depress keys "A," "B" and "X" on the keyboard, and not stop playback almost as soon as the ("mouse"--actually trackball in my case) pointer is moved off the "A," "B" and "X" buttons.
In df-metric there is no need to look inside a DUT as it is a black box.
In attempting to test the validity, or at least use, of "DF Metric" scores, what's going on inside the "DUT"--or at least some idea of its "transfer function"--would seem to be very much of concern.
A soft clipper would be expected to act more on percussive transients, etc.--in which case it's acting more on "inharmonic" components (that can cope with a lot of IMD--"smashing" snares is done in production, after all) and it would tend to make these perceptually louder, giving an increased sense of dynamics and energy. These transients would also tend to mask the "harmonic" content underneath, thus the effect of the clipper on "harmonic" components underneath too is masked. OTOH, the "Softube Trident A-Range" (with the setting used) would add some colouration to everything; it still modifies the signal below ~4dBFS, and, as shown, THD exists at considerably lower levels.
And that, BTW, is why I selected those processes in the first place--where the "error" may well be similar but the psychoacoustic effect, if audible, to all intents and purposes, must be different. I'm rather surprised that it turns out that the scores (that you obtained for them) are so very similar!
With the "DF Metric" results that I obtained for dithered vs. truncated versions of my reference file, the score was "worse" for the dithered versions; and when truncating to 4-bits, the signal is full of highly audible quantization distortion. So, that's another area that I think is worth exploring further.