Quote:
Originally Posted by Xenofan 29A
I do not consider this program any sort of be all and end all solution or explanation. I am entirely aware that it does not show THE ONE AND ONLY TRUTH of the matter. I am also aware that different programs will get different results.
|
In fact I'm getting the impression that it doesn't exactly model (humanly) perceived DR (see below).
Quote:
Originally Posted by Xenofan 29A
The only reason I brought it in was to illustrate a point that it was NOT a measurement of how good or bad something is.
|
Agreed. Especially how it seems to be computed as of today.
Quote:
Originally Posted by Xenofan 29A
I don't even think we're really disagreeing here.
|
Haha
-----------------
Anyway, I did a bit more digging and stumbled upon these:
Pleasurize Music Foundation @ HA
This at least partially reveals how the computation works. The bulk of information (sparse!) comes from the (german) manual:
Quote:
Zur Ermittlung des offiziellen DR‐Wertes wird der Titel bzw.
das Image des Tonträgers (Wave, 16bit, 44,1 kHz) gescannt
und im Hintergrund ein Histogramm (Lautheitsverteilungs‐
Diagramm) mit einer Auflösung in 0,01 dB‐Schritten erzeugt.
Die nach etablierten Standards für die RMS‐Berechnung in
einem Zeitfenster von 3 Sekunden ermittelten
Lautheitswerte (dB/RMS) werden quasi in 10.000
unterschiedliche Schubladen aufgeteilt (der 0,01dB‐
Auflösung entsprechend). Von dem Ergebnis werden nun die
lautesten 20% als Berechnungsgrundlage für die
durchschnittliche Lautheit der lauten Passagen errechnet.
Gleichzeitig wird der höchste Peakwert ermittelt.
Der DR‐Wert ist die Differenz zwischen Peak und Top‐20 des
durchschnittlichen RMS.
|
My problem with this: It doesn't take human loudness perception into account at any point. ReplayGain e.g. does this (see the aforementioned paragraph on HA about loudness computation), and also R 128 -- it should be mandatory by now. If you want a good metric for DR you have to consider the "consumer" of the audio, and this is (with high probability) going to be a human being.
So if anything this DR value is going to give you a _very_ crude idea. In fact the effects that you mentioned Xenofan ("bad" DR values but still nicely sounding) might be largely because of this issue (not taking into account the human aural system).
Another thread, this time on "JusticeForAudio.org":
Link
This brought me to this github repo by adiblol, who implemented the algorithm described in the manual:
dr_meter@github
He claims that the results from this tool are "compatible" with the TT DR Offline Meter (which is also the basis for the DR values in the database). No way I can confirm this though, but at least there is some "reference" code for what this tool is supposedly doing.
So what does all this tell us? At least for me it tells me that this DR values should be regarded with great skepticism. Like always, take them with a grain of salt -- preferrably a _big_ grain in this case.