(more) Audio Summing Malarchy

[pipelineaudio]

I deal with this in odd snippets here or there, but today I wasted a whole day, on nothing but this nonsense.

For some background, among all the audiophile nonsense, past the magic cable pixie dust, and into the world of the so called professional engineer, we have "the DAW summing Myth" that comes up in audio forums, stores, studios, and what have you, every 15 or so minutes

The myth goes like this: Mixing audio in some DAW's (digital audio workstations) adds super faerie dust magic to the dound and results in a hit. Audio mixed in other DAW's not currently featured in the industry's main magazine produces inferior audio that cannot be used professionally

Might seem reasonable, but in truth, each and every app is designed to be as transparent as possible.

garbage in/ Garbage out

WYSIWYG

there are some gotchas, dealing with the amount of bits you have to work with, but when it comes down to it, its just math. 2+2 BETTER = 4 or else

And normally it does.

Even if it DIDNT, I submit there is nothing, nada, zilcho, that could possibly ALWAYS be better. No matter what sort of distortion, be it tone, dynamics, time tricks, it doesnt matter. I submit NOTHING will *always* be better than transparent

But still it goes on.

Today, I blew the whole day making and soliciting the help for, some real summing tests, as follows.

****************

Being such a tired and ridiculous topic that seems to come up every 11 minutes or so, I wanted to be able to point to a link where not only would it be proven that this whole summing issue is nonsense, but where you could prove it to YOURSELF!

This actually wasn't as easy as I had imagined it would be, and there are still a few little kinks to work out.

Here are the test files to download, along with the results:

http://pipelineaudio.net/samples/summing%20test.zip

Here is how it works:

For the "Straight Test"

1. Place all three files from the "test files" folder in your project
2. Set project for 44khz
3. Turn dither off if possible
4. Render to 24 bit 44khz wave file





For the "faders down 6 master up 6 Test"

1. Place all three files from the "test files" folder in your project
2. Set project for 44khz
3. Turn all three channel faders down to -6dB
4. Set master to +6dB
5. Turn dither off if possible
6. Render to 24 bit 44khz wave file



Now, assuming that an app isn't BROKEN, every single one of these files should be identical

err wait, some of the tinfoil hat crowd claims some apps have magic pixie dust summing math that makes the sound have more SOAR, so I included the outputs just in case there were sonic differences, and they could claim they abx tested the files and found one "better" than another - yes and Im waiting for a certain Nigerian Ambassador to transfer funds into my bank account, since you know, I sent him the transfer fee.....


So, so far as you shall see, in the unity gain test we have:

REAPER, Podium, Vegas, Pro Tools, Cubase, Nuendo, SAW, Samplitude, SONAR and Logic 8

You already know the answer: Samplitude and SAW have magical supermixer properties that will make hit producing euphonanity over all these other apps right?

Wrong, every one of these is a DEAD, MINUS INFINITY, null

Given the same input, all these apps produce the same output...they sound EXACTLY the same. Unless you can posit a theory of how identical files could sound different, I think its time to start worrying about mic positioning, control room accoustics, or hey, the talent of the "talent", you know, things that might actually matter

But wait, there's more....just in case some of these apps might be doing things a little differently, we did a test of turning down the channel fader by 6dB, turning UP the master fader by 6dB and rendering. I mean, If there IS any magic pixie dust sprinkled in the code somewhere, its gotta show in the level changes, where else could it be?

Unless the app is somehow screwed up, it should still null with the unity gain files right?

well....here's where things got a little interesting

We KNOW 100% that pan laws arent an issue here, or the unity gain renders would not have nulled

so, Vegas, SONAR, REAPER, and Pro Tools nulled with the unity gain files exactly as they should.

SAW did not - what seems to be a slight error in SAW's math resulted in a render which was 0.00451dB too quiet. When this gain was added back in, the file nulled as expected - note to the Shakti Stone arrangers out there, it still didnt make a different sounding file, just a quieter one that sounded exactly the same as the others once the level error was corrected

Logic 8 also, did not render correctly, it resulted in a file which was 0.06dB too loud. Again, turning down the file to the correct level resulted in a file which nulled to -infinity with the rest of the files.

Cubase/Nuendo - here is the weirdest one. These files did not null with the others. The level was correct, we know the pan law was correct, I just dont have an explanation for why these didn't render correctly.

(actually I do, as was pointed out, this is a known issue with cubendo: http://forum.nuendo.com/phpbb2/viewtopic.php?t=13345&postdays=0&postorder=asc&highlight=null+summing&start=0 )

But before you send that check to Sylvia Browne, these nulled to -99dB, HARDLY anything to write home about






Now, it is QUITE possible there was some pilot error involved so I heartily invite you to download the tests yourself, and have at it.

For all practical purposes, I consider this summing nonsense dead and buried.

***********************************

So, being a skeptics board, do you see anything wrong with my methodology?

And no, this didn't put it to rest, not 3 minutes had gone by and Ive already started hearing woo theories of cumulative magic poofter audio

[graphicsguy]

Wow...I'm really glad to see some "technology skeptics" on here. I'm certainly not familiar with this "summing" thing...the closest I get to a DAW these days is Cakewalk's Project 5 and Adobe's Soundbooth...

I am amazed by the things that tech-geeks will tout along. It really doesn't surprise me it's entered the "woo" realm though.

[pipelineaudio]

BAH, as expected, the woowoos pulled the audio version of "you dont believe in god, why do you waste time arguing about him" argument on me

I am of course, very wrong in my findings, but noone will post a counter to it, just insult me instead.

I guess you guys saw that coming from ten miles off lol



p.s. graphicsguy, cakewalk's products (sonar, etc) summ fine :)

[ktesibios]

So, being a skeptics board, do you see anything wrong with my methodology?

And no, this didn't put it to rest, not 3 minutes had gone by and Ive already started hearing woo theories of cumulative magic poofter audio

I can see that you're starting with the principle that in the time domain, summing is simply addition, that is, A(t) + B(t) = C(t) and that gain/attenuation is simply multiplication by a constant, that is, A(t)*k = b(t) and seeking to test how well that is implemented in various DAW packages. So far, so good. Unfortunately, I can't find a description of your actual methodology in what you've linked, so there's no way to tell if there's anything wrong. In particular I'd be curious about the method you're using to compare the various output files. I remember that back in the day dear old SDII had a "file compare" function which would subtract one file from another, sample by sample and create a third file containing the difference, which could be a handy troubleshooting tool.

Being more involved with analog hardware these days than with mixing "in the box" I'm not familiar with what similar tools are available in current workstations, and I can't go play with the ProTools rig right now because there's a session using it... As an experiment I tried using the "invert" and "mix" functions in Bias Peak on this computer and found that doing this with any of your test files produced a null to +- 1 LSB, so that might be a valid means of comparison.

If the object is to mix the test files using different DAWs and compare the results, that's a valid way to smoke out differences in how they handle the math, but it doesn't tell anything about the absolute accuracy of any individual DAW. To do that I would proceed thus:

1. Generate a test signal file.
2. Duplicate the file.
3. Import both into DAW, sum them and save output to a third file.
4. If the DAW is perfect, the output should be identical to the input file, but with twice the amplitude, therefore
5. Subtracting the original source signal from the output signal twice should give a zero result: a file in which all sample values are zero.

To test the accuracy of the gain/attenuation, I would use a single source file, set the channel fader to -6 dB and the master fader to +6 dB. The result should be identical to the original, so subtracting the source from the output should give a zero result.

What's funny- to someone who started out working with analog tape nigh twenty years ago- is that the error nits which are being picked are so tiny. If you did a set of comparative freq response sweeps on, say, an Otari MTR-90 and a Sony APR-24, the measurable differences were orders of magnitude greater than any differences you're likely to measure between two different workstation software packages. Now there was a situation where engineer preference for the "sound" of one machine over another had a clear basis in reality.

BTW, in the environment in which I work (two fairly "high-end" studios), hardly any mixing is done "in the box". The clients generally use the workstations as multitrack recorders and do the summing in an analog console (where there are genuine, readily measurable differences between what comes out of the summing buss on, say, a Neve 8078 and an SSL 9000K).

There are also outboard analog summing boxes on the market for people who don't want to trust the DSP in their workstation.

De gustibus non est disputandum and the never-ending search for a magic pixie dust box (or as an old head back at Sigma put it, "snake oil of the month") goes on...

[pipelineaudio]

heres how it works:

I have included three test files. a 250hz sine, a 1khz sine and a sine that sweeps

you render them in your daw and then we put all the renders into another daw and invert the polarity of one, playing back two at a time. Im looking for them to null to -infinity, though honeslty, anything under -144dBFS is too low to hear on a theoretical 24 bit converter

[Fnord]

Well done! I can see the wooist response now...

I can hear the difference! So what if YOU have a tin ear? Your fancy equipment doesn't mean anything; some effects just can't be measured! Can you prove that I CAN'T hear the difference?

... I just hope that there is no JREF member named "AudioWooster" to take offense.

[ktesibios]

heres how it works:

I have included three test files. a 250hz sine, a 1khz sine and a sine that sweeps

you render them in your daw and then we put all the renders into another daw and invert the polarity of one, playing back two at a time. Im looking for them to null to -infinity, though honeslty, anything under -144dBFS is too low to hear on a theoretical 24 bit converter

I could expect a perfect null for a test of summing, which is simple addition and doesn't change the precision needed to perfectly represent the result provided you avoid overranging). In the case where you're testing the accuracy of gain change, you're multiplying, and multiplying an m-bit value by an n-bit value produces a result having n+m bits. This has to be re-quantized to fit the resolution of the output; whether you simply truncate or dither it down you're going to incur the same +-1LSB uncertainty exhibited by A/D conversion.

A null to +-1LSB should be taken as "perfect', just as a null to 0.

Meanwhile, there's a new professional analog tape just out and I'm getting my hands on a reel to run some measurements on on Thursday. Ohboyohboyohboy...

[pipelineaudio]

Right, even with the gain change, most apps are still nulling to -infinity

Its all in the topic post

[marting]

6dB is often used as a shorthand for doubling a signal level. It's actually about half a percent less. This would introduce rounding/truncation errors.

[pipelineaudio]

6dB is often used as a shorthand for doubling a signal level. It's actually about half a percent less. This would introduce rounding/truncation errors.

That is what probably caused some of the screwups, treating 6.02 as 6

However quite a few worked in WYSIWYG fashion and summed correctly