0.9u-5 - what's happened to the imaging when upsampling?

[Leo]

If it is 90 degrees off, than DSP can deal with it . At least that is a function on the Behringer DCX (and DEQ I think).

Well Peter, there you go, DSP corrections for double and quad......

[PeterSt]

If it's 90 degrees of, it's 50% wrong.

The kernel code producing the sound is some 6000 lines of program code now, which kind of makes you dizzy about the bits and pieces looking alike but are not. You could say the "smartness" in there to provide the good sound on one hand and save cpu cycles on the other, starts to cripple me a bit. Also, checking by listening has become a sheer impossibility because of all the combinations, and actually I am very glad that Mani "found" this means of checking.

Btw, I briefly looked at the aliasing, and things happen only at the other side (above 22050Hz) of the mirror. Only if you think you can hear this or you think it influences, you could call it wrong. Please note that no AA is applied on purpose, that leaving the audible data, say, bit perfect (and otherwise you'd get a roll off in the highs and more side effects).
The "hum" your picture showed in the other topic          , luckily didn't show with me (because *that* would really have been aliases at the wrong side of the mirror).
Careful : this was briefly looking.

[manisandher]

... this seems a very easy test for the quality of oversampling, whereas all kinds of semi-religious wars are being fought on the fora about which up-sampling method is better (software, (and then which...),  hardware, realtime or not) .

I would bet my house that there is more going on than a test like this can reveal.

For my own part, I believe so much is down to phase coherence. I'm just not sure what happens when aliasing etc comes into play. For example, in the quad upsampling case, the left and right are not 90 degrees out of phase... but some of the components are - I suspect those that arise from aliasing (though perhaps these should be 180 degrees out of phase?).

The situation is too complex for me to understand fully, but I do find these sorts of tests to be useful.

Will try get some images using upsampling in Foobar for comparison.

Mani.

[manisandher]

... I am very glad that Mani "found" this means of checking.

My pleasure.

Mani.

[manisandher]

Btw, I briefly looked at the aliasing, and things happen only at the other side (above 22050Hz) of the mirror. Only if you think you can hear this or you think it influences, you could call it wrong.

Alaising definitely influences the sound below 22.05KHz. It has a filtering effect on the signal as it approaches the 22.05KHz cutoff. That's why there is a lowering in the RMS level of the 15KHz signal in the quad upsampled image.

And perhaps that explains why I hear a softening in the sound and a lowering of transient attack when using quad upsampling? In any event, I think we can all hear some change in the sound, no?

Mani.

[PeterSt]

For example, in the quad upsampling case, the left and right are not 90 degrees out of phase... but some of the components are - I suspect those that arise from aliasing (though perhaps these should be 180 degrees out of phase?).

Ah, I'm sorry, I wanted to respond to that in the previous post, but got stuck in the 6000 lines of code.

Indeed, what will have happened, is that PARTS of the samples (oh yes) have been assigned to the wrong channel, or just aren't upsampled (but repeated = doubled (or quadrupuled)). This may look strange as something which can happen (in decent program code), but it really is not if you know what all has to happen to utilize the additionally available bits (because remember, this is so far about 16 bits only, while 24 are avaialble, and I just use them). All in combination with the digital volume, DACs which allow for 16 bits only *but* where Doubling/Quadding is allowed just the same, and some more stuff like the calculations themselves which may be just wrong.

Anyway Mani, what I see here indeed, is or :
a. a mixture of aliasing and wrongly assigned sample parts or
b. all upsampling just being plain wrong (on that matter I forgot to check "DAC is 16 bits") or
c. aliasing influencing the whole spectrum, but which doesn't show on the frequency plots.

At this moment I can't even see how a. can be in order, but I keep in my mind that the RME software just *will* show the theoretical frequency full range, coming from the sampling frequency, hence up to 44100 Hz at 88K2 sampling. So it will incorporate the anomalies above 22050 just as well (but you can't see them seprately in the Gonio plot I think).

But let me first try to find the now 100% sure culprit in the program. Then we can start looking for other things.

[PeterSt]

OK. I've got a small "confession" to make;

Double/Upsample is working OK now here, and I imagine myself to have heard the most undisturbed sound so far. Don't matter higher buffer settings and all, I enjoyed the music VERY much the last couple of hours. Does it work afterall then (my remark about movies where upsampling just visually works for 100 % sure) ?

Quad/Upsample does not work yet other than "wrongly", and it kind of needs the program to be re-written which I will do.

Mani, thank you once more for your remarkable justification on ... well, the title of this topic;
With or without justifying pictures I listen. Now or in the future.

Peter

[PeterSt]

What happened to the imaging ? Well, I can tell you.

Quad/Upsample :
First pair (L/R) of samples (out of the total of 4 pairs) was written ok, then 3 interpolated samples went to the left channel, then 3 interpolated samples went to the right channel. So, 25 % of samples was ok, 75% ended up in the wrong channel, implying a high degree of mono sound.

Double/Upsample :
3 wrong lines of code indeed caused that a.o. the part of one sample (think of first 8 bits vs. the last 4 bits) ended up in the wrong channel.
This was expected as per the plot Mani showed (a torned circle).

It's actually fun learning to read these plots related to program errors.
Again, Mani, thank you.
Peter

[manisandher]

My pleasure.

Yeah, the genios are fun, aren't they?

I suppose I can speak for all of us; can't wait to get hold of 0.9u-6 

Mani.

[LydMekk]

Yep, me too. Have gotten so used to DoubleOn, UpsampOn that going back to vanilla 44/16 has been quite a chore...

Awaiting version 6 with anticipation.

[Leo]

quad oversampling sounds great with number 6 !