Hi, ALL,

New objective audio parameter is proposed. The article is here - SoundExpert
Your opinions and criticism would be helpful.

Serge.

How does this method compare with PEAQ?

ff123

I'm unsure of the goal. In the "Limitations" part you clearly state that this measure isn't usefull to determine perceived quality.

Also, if you look at the four graphs in that chapter, on the second sample we see that the D_f score of WMA and Vorbis is about 20dB apart while they score very similar in the listening test (considering the probability margins, they have the same quality). If I understand correctly, you are comparing power and 20dB would then correspond to a factor 100!

Said differently, this measure can deviate by a factor of 100 without change in perceived quality.

In chapter four a possible other use is discussed:

I doubt this is practical.
For this to work you need to have both an original and a processed version.
As long as you have the original you can run it through any encoder you like and compare the result to the processed version which should be much more accurate.

Granted, this method requires constant access to many codecs and in the long run requires more processing power. But judging by the variance of D_f value for a specific codec it will be difficult to reach a satisfactory confidence level.

I'm sorry this turned out to be quite destructive criticism.

Well, it’s hard to compare one very “instrumental” parameter with the entire approach towards estimation of audio equipment quality like PEAQ. In short, it’s just an element of another approach, which is a little bit different from PEAQ and completely independent of the latter. I hope to show the difference in the next paper.

It’s useful in case of amplifiers, cables, processing in audio editors and so on. There it has strong correlation with perceived quality and could play the role of “advanced” THD. Psychoacoustic coding is more complicated case, in which D_f is useless as any other objective parameter, but could be useful as a part of more complicated measurement methods…

It’s surprising to the same extent as the fact that 10% of sound file could sound almost like 100% (~140 kbit/s). It’s more the miracle of human hearing than the drawback of D_f.

What exactly you should compare in two signals (files)? I propose to compare the shape of the signals with the instrument – D_f. But I could agree with you that this example is more illustrative than practical.

One problem with this approach (apart from the obvious wrt psychoacoustic coding) is that, typically, we want the difference to be zero. That's (hopefully) a trivial thing to check in the digital domain.

Where the difference isn't zero, we're simply too clever these days for these older measurements to be useful. Take an audio editor converting 24-bit data to 16-bit. The measured difference is tiny if we just truncate, but huge if we use strongly noise-shaped dither. Which is better?

It can be argued that the approach is more suited to analogue devices, e.g. amplifiers. Here again, it's quite easy to make a device that measures better (in this respect), but sounds worse. You could argue that any measurement can have this problem. I'd answer that this is true, but it doesn't necessarily mean that we need another one!

However, this is all being a little unfair. I think we've all used input-output difference analysis (I-O D A - that's the only name I could find for it in my PhD) for something some time, with or without averaging, summing, psychoacoustic models or assumptions, spectral or temporal analysis etc. So it clearly is useful, and sensible people know its limits. Simply summing the power of the difference gives a nice single number, but probably hides much of the useful information that comes from examining the difference. One big advantage of I-O D A is that you can use real signals, rather than a restricted set of test signals.

One problem with input-output difference analysis is that it works best in the digital domain. With analogue equipment, if you are going to genuinely perform this analysis, you need to use test equipment which is better than what you are testing. This is true with all measurements, but normally you have a very good idea of what you are measuring, so the test equipment only has to be good enough in this or that domain or quantity. Since you are testing everything with I-O D A, then the test equipment must be better than the system under test in every possible respect. This is a tall order - not always impossible, but not easy to be certain of either.

I hope these thoughts are useful.

Cheers,
David.

Actually, I am agreeing with all said above. Nevertheless I am sure that some kind of basic integral parameter for measuring difference will be claimed forever as a part of the following approach – first, let’s see the amount of a difference, then find out the perceptual significance of that difference. The more so as the first part of this dichotomy is very simple in measurement. On the other hand, for a long time THD (with all its terrific drawbacks) was the most important parameter for audio engineers, but there were made a lot of really good equipment still. D_f, in these terms, is more indicative as it allows using more ‘real life’ signals than sinus.
Can you point out some other integral parameters widely known and adopted for measuring difference?

So the aim of the article was to show that some kind of pure instrumental parameter is and will be necessary and Difference level is very very suitable one.

Doesn't every measurement in existence aim to measure the difference between the wanted (ideal) output, and the actual output? Maybe this is a generous definition of difference.

Sorry if that's not a very helpful response, but I'm not entirely sure I understood the question!

Cheers,
David.

Yes, I was not quite specific. Of course I meant difference signal. It is widely used in various research works de facto and I made an attempt to define it de jure by introducing the parameter, which measures the quantity of that difference signal.