16/44.1 vs. 24/176.4

[PeterSt]

I really have no idea. I think I can only know when I really measured. One way or the other. Also see next posts ...

[PeterSt]

Ok. With my own statement in mind "an AI upsampled file can't be compared", I thought "but wait, this never was an AI upsampled file ... this was the result of an OS DAC with something "like that" in it. So, let's see ...

Was I wrong on thinking AI couldn't be compared ...
What you see here is just that; At the top is the original 176.4, and below it you see the AI upsampled file, with Josef's 16/44.1 as the base.
Remember, the 16/44.1 was the decimated version from the same 176.4 you see here at the top ...

Yes, in the top channel (at 1/3, but notice the scrollbar at the bottom !) you see an anomaly, the AI version showing a bump, the original just showing a slightly upwards going slope.
Notice that I can not know whether this is cause by my AI upsampling, or whether it is caused by the decimation process from Josef. And no, I am not going to find this position in his file.
But these anomalies are very very rare; I have been browsing through this step by step for 45 minutes, and found only two of them (see later for the other one). Can this change sound ? maybe, if only enough are there. My 45 minutes of browsing was worth 6 seconds of music ... (I started at 1 minute, going backwards).

Also see next posts for some clues ...

[PeterSt]

A little zoomed out. What we see here, I only saw after a long time as a pattern ... what do you see ?

Look at each of the peaks and compare the both takes ...

Do you see it ?

Somehow there is a consistent "opposite" balance in how the peaks are formed. Just find a peak where one sample bumps out. It doesn't matter where you find it. Now look at its counterpart. See ? the one with one sample jumping out has 3 samples to *let* that sample jump out, where it counterpart has an even number of samples for the width concerned, and it is impossible for one sample to jump out.

If you try to reason where the "base" of this emerges, at least it is hard to see, but it must be at the base of the wave somewhere, where the one is wider as the other.

This is a most consistent thing throughout the both files !

Will this change sound ? ha, you bet it will.

[PeterSt]

Two other examples, and nothing new. The first one contains an anomaly again (at 2/3, but notice the scrollbar at the bottom !), and further you again see what I talked about in the before post.

Notice that it is not in 100% of cases like the rightmost two pointing down ones (the second one is just visible in the upper right corner of the bottom half (screenshot) of the first picture).

But generally this upsampling means is very very much alike the original. What ? it is 1000 times better than I ever expected.
I should even listen to it ! haha

Peter

EDIT, PS: I'm not sure whether it can be the cause, but it looks like the upsampled version is one sample shifted opposed to the original (like skipping the first sample). In the end I don't think this is true, but if you look at the first post of this last sequence, you can see that the bottom part is not time alined with the top part. Still it is ... So, if the upsampler shifts one sample, then you'd have something like this. I think.

EDIT2, PPS: I just saw that the 2nd picture in this post, at the very right (use the scrollbar for that) also contains an anomaly. I didn't see it at first, so there may be more than I have noticed.

[manisandher]

I should even listen to it ! haha.

QAI applied to my native 16/44.1 came 2nd in my 'Objective Assessment', after the native 24/176.4. I'm going to give QAP and QAI applied to Josef's 16/44.1 file a good listen... might be interesting.

Mani.

[manisandher]

Yes, in the top channel (at 1/3, but notice the scrollbar at the bottom !) you see an anomaly, the AI version showing a bump, the original just showing a slightly upwards going slope.
Notice that I can not know whether this is cause by my AI upsampling, or whether it is caused by the decimation process from Josef.

If you could repeat QAI with 'my' 16/44.1, you could see if these bumps are there also. If so, it's likely to be AI doing this, no?

EDIT: No, of course you can't - they're not aligned, and you're not going to do that, right?

Mani.

[manisandher]

Apart from the odd anomaly, and the strange behaviour at the peaks, these graphs look remarkably alike, don't they? Yeah, 1000 times better than one would expect from AI!

Now, is there a way to speed the processing of AI up???

Mani.

[PeterSt]

No, of course you can't - they're not aligned, and you're not going to do that, right?

If I had to look for that same spot ? undoable. This merely has to go like seeing a recognizeable point (a strange shape when zoomed out) and find that in both files by slowly zooming more and more. Can take 30 minutes for such a recognizeable spot. But without such a spot (zoomed out) ? even if you had a day ...

Now, is there a way to speed the processing of AI up???

If there was that I could think of, I had applied it.
Yea, applying a more simple filter. Haha.

Peter

[Josef]

WOW! Great analysis Peter. Thanks.

But all sure tells me that Mani's 44.1 MUST sound different.

I agree with this 100%. What I call the native-16/44.1 sounds the most different from the vinyl. And the native-24/176.4 sounds the most similar to the vinyl (actually identical, as far as I can tell on my system). Josef's 16/44.1 definitely sounds more similar to the vinyl than my 16/44.1.

Mani.

Try this:
http://www.filemail.com/confirmation.aspx?id=HAHCNWRYCNBXSTE

I am curious if you find this version more similar to your 'native 16/44'. (it is also derived from your 24/176 master but in a different way)

If you do find it 'more similar' then we might have some interesting food for thought

[manisandher]

Yes, definitely 'more similar'... but still different...

Firstly, I like the sound of the your new 16/44.1 - it's got 'life'. I listened to it all the way through first off and really enjoyed it. However, it's not as close as your original 16/44.1 is to the 24/176.4. The main difference seems to be in a 'thinning' of the sound. Everything seems to sound clearer, but at the expense of body to the instruments. In this respect, your new 16/44.1 is remarkably similar to the effect that QAP has on my native-16/44.1.

If I had to choose only one 16/44.1 file, it would be either your new 16/44.1 or my native-16/44.1. BUT... not because they sound like the 24/176.4 - they don't. But rather because they have 'life' and are just so enjoyable to listen to. My one caveat with your new 16/44.1 is that on prolonged listening, it might get a bit fatiguing (again, like QAP on my native-16/44.1).

So please, let us in on it - what did you change? I'm intrigued...

Mani.

[PeterSt]

Maybe it is a little off topic, but after the "ages ago" I created the AI upsampling, I now judged it again.

Oh boy ...

I remember that after I created it I used it for two months or so, just because I was very satisfied with it. But did time proceed ...

AI TO_TAL_LY destroys the sound. It just makes complete rubbish of it. But, nothing much different from the Weiss Minerva I listened to the other day. So please keep in mind - it is just what one is used to (like me liking AI before, which was on the same NOS/Filterless DAC, though in a much much lesser (SQ) version of it).

But maybe you can squeeze out the same as I did;

Maybe you have Avishai Cohen - AT Home. Well, you should get it if you don't have it anyway.
I listened to track 04-Remembering;
This is a rather sensitive track, which first of all now doesn't "work" at all. But besides that all the piano notes seem cut, and there's nothing like a real bass around. But more interestingly, while the guy plays the bass in a to me strange fashion - like plucking two strings while the same note (frequency) comes out (can that be done with one very large left hand ?) - this interacts in AI in a way that additional harmonics emerge which make the man play totally off-key. It sounds interesting, but it shouldn't be there (later I checked with AP and there it's just two the same notes).

I played some ambient piece you won't have anyway, and I already learned from the Minerva that when things become the most fragile, it will bring the most distortion. And indeed it did. Nicely sweeping sounds just become shattering noise. Yes, I guess it is hard to believe, but it just is so. Funnily enough, when you are not used to how it should sound, you could accept it as normal, but then to the sense of "some lousy band playing a lousy synthesizer lousily".

I played "The Weight" from Chris Barber. It sounded the most normal of them all, but I noticed it was very loud. Actually so loud that the trumpets shattered a few tads too much to my likings. Later, with AP everything appeared to be just normal.

So ...
AI may then be able to follow the waves in an unexpected good manner, it is and remains a filter which should ring as hell. Why I can't find this in the data I don't know, but most probably that will be a matter of just not being able to see it in those kazillion samples. And now I think of it, I just heard a toing-toing-toing echo in one of the tracks with AI of which I'm sure I never heard it. But I forgot to check it back with AP, so not sure about that. What I do now is that the shattering distortion I talked about, *always* is something which repeats/smears. So, when there would be an original sound of say 1 second, when it is turned into a shattering noise this will be 3 seconds.

Am I glad I did this. Because, as it now turns out "I" am lightyears further than I knew myself. It just needs recalling the time I used AI for two months. How can I ever have done *that* ... It is unlistenable.

Spoiled me.

[Josef]

>So please, let us in on it - what did you change? I'm intrigued...

OK, here it comes but please be forewarned: you may not like what you'll hear so no shooting the messenger, OK?

Idea was this: I suspected that 'dynamic decimation' _was_ indeed a marketing term to just keep 1 out of 4 samples or some very, very simple (=poor quality) downsampling.
More importantly, I suspected that ADC also did some limited/poor quality dithering or no dithering at all i.e. it just chopped last 8 of 24 bits.

So if this theory was correct, by taking your master and doing the same thing to it we should be getting close to the sound of your native 16/44 version!

Now, I don't have software that can do _exactly_ that - perhaps somebody else does - but I was convinced if I just did something _similar_ we should be getting closer to your native 16/44 sound.
  
So, for step 1. what I did is to take the worst resampler I had at hand: Foobar's PPHS. PPHS still does much more than simply dropping 3/4 samples but is clearly inferior in quality to professional algorithm used for previous version. This is important as downsampling creates all kinds of side-effects and lower-quality algorithms create more of that (think it's called aliasing or quantization noise or somesuch .  BTW - I could not get as close to your native 16/44 sound by using high-quality resampler!

Then, I simply skipped steps 2. & 3. from first version (Noise Shaping & Dithering) and just 'cut' the bits to first 16. Note that this will add _even more_ quantization noise to result!

While this step also is not exactly the same as taking 24 bits and discarding last 8 it is close enough: strictly technically speaking when PPHS dowsampled from 176 to 44.1 it created 64-bit floats and then those 64-bit floats were just cut by Foobar to 16-bit integer in the simplest manner = without any processing.

So what does all this mean? Here's one explanation I believe is at least close to truth:

Initially, I agreed with your assesment of 'native 16/44': it  _seemed_ to have more 'buggy factor' or 'life' as you put it. But when I compared it with my manually downsampled version (note that some amazing software was used as I explained in email!) I suddenly found it too 'rough'.
Then I remembered where I heard that pattern before: as mentioned in my post, some time ago when I was processing files for Ipod (=applying Replay Gain when converting to WAVs), I turned off Noise Shaping & Dithering to experiment (you can try the same!)
And the track I used also had a leading piano And it also initially sounded like it had 'more life' and I thought I heard the piano screeching & making all kinds of funny noises which I thought uncovered more detail & realism that previously was masked! So I converted bunch of music this way!
Long story short, after some time (and it wasn't minutes, it was probably days - I'm getting old, lol:)) I realized that I was getting fatigued after relatively short listening sessions - just like you found out in your latest post!

So, there wasn't more 'realism' or more 'life' there - it was just more quantization noise that fooled me back then and, if I may respectfully suspect, you too this time:)

Cheers,
Josef

[PeterSt]

Hahaha ... I'm not sure what the conclusion is now ... but listen to both your versions Josef, and Mani's (the 44.1) ... about quantization noise ...

A few minutes ago I already typed "but the quantization noise should be audible ?" ... and then I thought to listen for it. Josef, your both versions are full of it. Mani's has NONE.

You can hear it right in the beginning, when the needle is still in the lead-in.

Sorry ... (but maybe it was intentional ? ... I can't be sure readig your last post)

(and besides, Mani's version keeps on sounding the best to me)

Peter

[manisandher]

... you may not like what you'll hear so no shooting the messenger

No, no chance of this. I'm always happy to learn...

I suspected that 'dynamic decimation' _was_ indeed a marketing term to just keep 1 out of 4 samples or some very, very simple (=poor quality) downsampling.

There is one thing that I'm absolutely certain of - the Pacific Microsonics Model Two does NOT have poor quality downsampling capabilities. I know of at least one highly respected mastering engineer who prefers its downsampler to ANY software downsampler - and he's tried them all, including the one you use Josef.

More importantly, I suspected that ADC also did some limited/poor quality dithering or no dithering at all i.e. it just chopped last 8 of 24 bits.

This is not the case. The Model Two provides a number of 16-bit dither options, which I understand are pretty sophisticated. You may or may not be aware that all HDCD-encoded CDs are mastered on the Model One/Two and have all been decimated down from its internal resolution of 24/176.4 or 24/192. In any event, I haven't come across an HDCD-encoded CD that doesn't sound anything other than superb.

So, there wasn't more 'realism' or more 'life' there - it was just more quantization noise that fooled me back then and, if I may respectfully suspect, you too this time:)

Maybe. But remember that my strong preference is for the analogue vinyl and the native-24/176.4 file, neither of which has this issue.

So, how is it that I can live with my native-16/44.1 and your new 16/44.1, even though they clearly sound very different from the vinyl and the native-24/176.4? (And by the way, don't get the impression that they sound the same as each other, as they really don't.)

If I had to hazard a guess, I'd say that I generally prefer as little DSP as possible. The more there is, the more the 'life' seems to get sucked out of the music. For example, I can't listen to your original 16/44.1 - I simply switch off.

I'm not sure where this leaves us really. But I'll just reiterate that I firmly believe that 16/44.1, although not perfect, can be made to sound really enjoyable.

Mani.

[manisandher]

AI TO_TAL_LY destroys the sound.

I've just tried it again, and agree with you (and with my initial 'Subjective Assessment') - AI kills the sound. (It also takes way too long to load.)

Mani.