Integer is King? - the challenge

[Chris Walton]

As you can see, it stores the floating point value in memory in 32 bits. It always reports a delta of 0.

Of course
32bit ieee floats can represent 24 bit integer values perfectly; the opposite is not true and you will need some kind of dithering (possibly with a valid psycho-acoustical correction model - maybe with feedback error correction too) to convert from float32 to int24. You will lose information anyway.

I know this is nitpicking but there is one int24 value that float32 can't represent - -2^23 (nitpicking, i know)

[jens]

You didn't ever hear a difference when there wasn't one? As in you clicked on "new take" or whatever it's called in your ABX tool, and it actually didn't change the algorithm, and you noticed a difference anyway?

It still all maybe coincidence and trust me I really hope it's all coincidence.

sorry that I have to chime in here but it's possible to make sure it's not coincidence (i.e. random) - you need to take a certain number (has be calculated) of samples (trials in this case) and once the number of deviations from the arithmetic middle passes a certain threshold (has to be calculated) the result can not be random anymore and thus becomes scientifically significant. Below this threshold the result is just coincidence. Quantitative science has very strict rules which are based on stochastics.

[gold]

Remember Clarke's 3rd law -- "Any sufficiently advanced technology is indistinguishable from magic."

Any thread with an Arthur C. Clarke quote is indeed the best thread!

[Samb]

sorry that I have to chime in here but it's possible to make sure it's not coincidence (i.e. random) - you need to take a certain number (has be calculated) of samples (trials in this case) and once the number of deviations from the arithmetic middle passes a certain threshold (has to be calculated) the result can not be random anymore and thus becomes scientifically significant. Below this threshold the result is just coincidence. Quantitative science has very strict rules which are based on stochastics.

Sorry, but it is not possible to be sure that something is not due to coincidence using statistical methods, only that it is unlikely to a specified degree.

[Aldo]

As you can see, it stores the floating point value in memory in 32 bits. It always reports a delta of 0.

Of course
32bit ieee floats can represent 24 bit integer values perfectly; the opposite is not true and you will need some kind of dithering (possibly with a valid psycho-acoustical correction model - maybe with feedback error correction too) to convert from float32 to int24. You will lose information anyway.

I know this is nitpicking but there is one int24 value that float32 can't represent - -2^23 (nitpicking, i know)

The range for 24 bit ints is from -(2^23) to (2^23)-1 right?
You have only 23 bits in the mantissa, but you also have an exponent!
If you turn -(2^23) in floating point 32 by hand you will find out that:
mantissa = 0 (implicit 1)
exponent = 23 (read 2 raised to +23)
sign = -1

Multiply all the values and you will get a -(2^23)

Am I missing something?

[vonRed]

I know this is nitpicking but there is one int24 value that float32 can't represent - -2^23

Are you quite sure about that?

Edit:

OK, I was slow again.

Am I missing something?

No.

[anti--hero]

i dont have a clue what half you guys are talking about but i get the general gist and very interesting none the less

have u ever tweaked a plugin becoming more & more perked and wide eyed with every change thinking to yourself by god i;ve done it, this sounds much punchier , heavier , sweeter, tighter...

..only to discover when u go to freeze the track it wasnt even turned on...

or am i just a friggin idiot ? (dont answer that )

i think i am - 9 pages and i still not sure what the outcome is

is there a diffrence or not ? or is the diffrence in an inaudible range therefore irrelevant ?

[whyterabbyt]

BTW, the 480l is identifiable by the way resonances occur (it sounds smoother), the reverb blends with the background, the overall sound is more one of space rather than a reverb tail and the way transients are handled (convolution smears). But Altiverb is really not bad, much better than many other plugins.

...and most of the people in the gearslutz thread preferred the altiverb. funny ol' world aint it.

[jens]

sorry that I have to chime in here but it's possible to make sure it's not coincidence (i.e. random) - you need to take a certain number (has be calculated) of samples (trials in this case) and once the number of deviations from the arithmetic middle passes a certain threshold (has to be calculated) the result can not be random anymore and thus becomes scientifically significant. Below this threshold the result is just coincidence. Quantitative science has very strict rules which are based on stochastics.

Sorry, but it is not possible to be sure that something is not due to coincidence using statistical methods, only that it is unlikely to a specified degree.

yes, you are correct but this specific degree goes towards zero (i.e. closely enough to conclude significance) under those premises I mentioned - Chebychev's inequality dictates this. even if each single result can be random the great number of samples will make almost sure that the total result won't be.

http://en.wikipedia.org/wiki/Chebyshev's_inequality


I hope I got this right - I has been about eight years since I've had my statistic exams...

[stefancrs]

You didn't ever hear a difference when there wasn't one? As in you clicked on "new take" or whatever it's called in your ABX tool, and it actually didn't change the algorithm, and you noticed a difference anyway?

It still all maybe coincidence and trust me I really hope it's all coincidence. That would make everything much easier. If I would have heard 15 times out of 15 times a difference than I would say that I really can heard a difference.
I'Ve done the ABX test prior to any further SNR analysis. If I'd repeat it now, I'd surely would loose myself in testing. It was just the kind: 'Let's see, what happens if I do an ABX test'. I didn't expect anything and now I better stop saying anything.

That's not what I was wondering. How many "new takes" did you make? Basically, what I'm wondering is if some of the "new takes" where fakes where the algorithm didn't change. The reason I'm wondering is just because I want to know if there where "blinds" involved in the test or not, or if the algorithm always were swapped for one of the others. Maybe I'm just daft though

living sounds: if you really are after better sound quality you first have to properly define what it is that you're after in one specific hardware. Saying things like "clarity", "punch" or "definition" means nothing, since they're all very fuzzy and very subjective descriptions of how something sounds. And it's obviously not even a "integer vs float" discussion, because there ARE software effects and synths using integer implementations, and yet you mean that hardware is superior. Maybe it's just the noise and distortion you're after?

[living sounds]

Some people probably preffered it because it is more audible as reverb (rather than adding depth perception). Lexicon's tendency to blend with the dry track makes the other one stand out a little in comparison.

In the end you always have to work with a tool in the studio to find out how useful it is. I certainly wouldn't give my Lexicon reverbs away for Altiverb nor any other plugin currently availible. But I still hope developers hurry to catch up quality-wise cause working with hardware is a PITA IMHO.

BTW, the 480l is identifiable by the way resonances occur (it sounds smoother), the reverb blends with the background, the overall sound is more one of space rather than a reverb tail and the way transients are handled (convolution smears). But Altiverb is really not bad, much better than many other plugins.

...and most of the people in the gearslutz thread preferred the altiverb. funny ol' world aint it.

[whyterabbyt]

Some people probably preffered it because it is more audible as reverb (rather than adding depth perception). Lexicon's tendency to blend with the dry track makes the other one stand out a little in comparison.

whatever makes you feel better.

[living sounds]

Maybe it is distortion. It's definitely an important part of the "record"-sound and of analog gear.

But with the purely digital stuff? It's very difficult to find an aedequate discription for complex non-visual sensory perceptions. Many plugins sound, for lack of a better word, blurred to me. Like lots of jitter. The frequencies are there, but impulses, transients get smeared. You can see this in an audio editor if you send for instance a kickdrum through the Lexicon300 and CSR with similar programs. Apply a low pass filter and see how the envelope of the sample that went through the plugin has lost contour, definition. It looks blurred. It certainly sounds blurred as well. The same goes for the transients in the mid and highs of course, but it's not as easy to visualize. The sample that went through the hardware has a much "faster" envelope. Convolution does this to audio as well.

living sounds: if you really are after better sound quality you first have to properly define what it is that you're after in one specific hardware. Saying things like "clarity", "punch" or "definition" means nothing, since they're all very fuzzy and very subjective descriptions of how something sounds. And it's obviously not even a "integer vs float" discussion, because there ARE software effects and synths using integer implementations, and yet you mean that hardware is superior. Maybe it's just the noise and distortion you're after?

[Christian Budde]

That's not what I was wondering. How many "new takes" did you make? Basically, what I'm wondering is if some of the "new takes" where fakes where the algorithm didn't change. The reason I'm wondering is just because I want to know if there where "blinds" involved in the test or not, or if the algorithm always were swapped for one of the others. Maybe I'm just daft though

Do you know how ABX tests work?

Here's an example A = 'Algorithm X', B = 'Algorithm Y', X = ???
You now have to say if X is A or B. If after 15 trials every guess was correct, you can assume that you've heard a difference. What 'blind' shall I put in there? X can only be either A or B.
After I finished the test I swear I had no idea about the result.
Ok, lets just ignore my lousy test and say that it all was coincidence, ok?

[Chris Walton]

my mistake

must be all the sleep I'm not getting.