Bouncing Down at Highest Quality Possible?

  • Thread starter Thread starter Steve Henningsgard
  • Start date Start date
Steve Henningsgard

Steve Henningsgard

New member
As I was bouncing down some mastering stems this evening (err.. morning, it would appear. Whoops.), it came to me that, while the session was tracked at 24/48, the plugins will run at whatever resolution they like. Which begs the question: if you track at 24/48, then import the tracks into a 24/96 session, will the plugins sound any better? Specifically, the ones creating sounds themselves (delays/reverbs, for instance).

Just a random thought...
 
dont see why it would, besides virtually anything over 44.1 is a marketing scheme.
 
It's been argued that non-linear processing (EQ, compression, limiting) can benefit by upsampling since it will spread out the upper frequencies and allow filtering above Nyquist to be handled better.

Some plug-ins and digital hardware (Weiss units) allow this as an option or do this automatically within their internal processing.

All of this assumes that you have a good SRC, otherwise you may be doing more damage.

A site worth checking out:
http://src.infinitewave.ca/
 
Last edited:
From my understanding, there wouldn't be any improvement. After you record at 48khz, there is already no high freq information above approx 24 khz. I don't think you can create high end freq information out of something that's not already there.

This calls for an experiment! Give me about an hour and hopefully I can come back with something useful :).

Eric
 
From my understanding, there wouldn't be any improvement. After you record at 48khz, there is already no high freq information above approx 24 khz. I don't think you can create high end freq information out of something that's not already there.

This calls for an experiment! Give me about an hour and hopefully I can come back with something useful :).

Eric

This is true, upsampling shouldn't add any additional information, non-linear processing may however. It's in the processing where there may be a benefit.
 
Okay. I tried to run my little experiment. The process went as follows.

I used a 44.1khz Nine Inch Nails song. Not sure of the name of the song but I have the file from a past audio experiment.

To view the frequency information, I used a program called Sonic Visualizer's Spectogram. This program is free for download if you're interested in visualizing your wave forms or frequency content. Great for LOOKING at your audio ... http://www.sonicvisualiser.org/

The first image I am posting represents the spectogram of the song. For those who aren't familiar with spectograms.. It is a visual representation of the frequencies that are being excited in the recording. The Y axis represents the frequency range and the X axis is time. This song is actually probably a poor choice for showing how the spectogram works. The frequency information is so dense because the mix is very very busy and extremely loud. You could usually see the dynamics and hits in the song.. this song is an exception .. but I digress.

In this image you can see that the frequency information stops at about 21 khz. This is because it is a 44.1khz cd quality song. Again, if you don't know already, the highest frequency a CD can reproduce is around 22khz. The highest this spectogram goes is 47khz. If this were recorded at 96khz with gear that functioned at those frequencies, the spectogram would have information all the way to the top of the Y axis.

When looking at this image you see that above 22khz are all void of content. This experiment is testing whether using non-linear signal processing can add higher frequency information when upsampling.

attachment.php


This second image is the second audio file. This file IS at 96khz. I imported the CD quality file into Pro Tools and added non-linear Waves Rverb, a reverb plug in capable of non-linear processing and 96khz sample rate (image of plug in below spectogram). With Rverb, I boosted all treble content as much as I could to see if, when upsampling, the content would affect any higher frequencies above the 22khz mark. (I also only used about 30 seconds of the song just for the experiment, so the clip isn't identical on the x axis but we are concerned with the Y axis in this experiment. )

attachment.php

attachment.php


The result was nil. There was still a brick wall (at 21khz) of high frequency content. This leads me to believe that upsampling does not affect higher frequences even when non-linear processing is applied.
 

Attachments

  • Original 44.1 NIN file.webp
    Original 44.1 NIN file.webp
    32.6 KB · Views: 175
  • Upsampled 96 khz NIN.webp
    Upsampled 96 khz NIN.webp
    44.7 KB · Views: 177
  • waves Rverb non linear.webp
    waves Rverb non linear.webp
    28 KB · Views: 170
Last edited:
Hi Erik,

I see that you put a great deal of work on this which is great but a couple of things. Firstly alias frequencies are frequencies which are created below Nyquist due to issues like improper filtering of frequencies that are created above Nyquist, not frequencies created due to non-linear processing which exist above.

See: http://zone.ni.com/reference/en-XX/help/371361B-01/lvanlsconcepts/aliasing/

These would be easier to see in an RTA if you had used a pure sine wave over a complex signal. If you have time I would suggest that you try taking a higher frequency sine wave and running it through a limiter so that it creates something closer to a square wave which will produce harmonics above Nyquist (unless the limiter is removing them as part of its algorithm) and try the same experiment. From there see if frequencies below the original frequency (alias frequencies) are created.

I would try here but the snow storm is keeping me away from the studio.

Best,
Tom
 
Last edited:
Erok,

A couple of doubts I have about your experiment. First is the idea - if I understand it correctly - is that the higher sample rate can affect the quality of processing and the frequency range in which it it can operate, but that does not automatically mean it's going to inject higher frequencies if the effect does not call for it.

In the case of reverb, I'm not sure that it's going to generate any higher frequencies from a source that didn't have them to begin with. In fact, usually as reverb decays, high frequencies actually tend to attenuate.

Now, I'm not 100% on all that, so maybe someone might be kind enough to correct me if I have something wrong, but if that holds, it would be quite understandable that you'd not get any higher response out of the test.

BTW, just as a side note; if you like those kind of time spectrum displays, you should check out the latest version of Adobe Audition. All versions of Audition have a nice spectral display tool, but the new version has turned it into an *editor* as well; i.e. you can actually edit/process your file by going in and clicking/painting/erasing/etc. directly on the spectral display. Very powerful and very cool! :)

G.
 
The "resolution" of a plug isn't really conected to a sample rate but the number of bits they use to do the math as far as I understand it. When you use such a plug it may increase the number of bits used in the math to theoreticaly reduce rounding errors in the math calculations (which may or may not be audible in the grand scheme of things)
But as other have said if you are upsampling the track the plug cannot create aditional samples that are not recorded in the original

I guess some plugs may upsample specifically to give more range in the nyquist frequencies so they can use a shallower high pass filter to reduce the chance of audible artifacts as a side effect of the processing but I don't know what they do in the case of the track providing 48k samples per second but the plug processing 96k samples per second. Process empty space to gain a shallower HP filter? that's a question for the math geniuses of the world but it would suggest that such a plug would add at least 50% more CPU load than running at the actual recorded sample rate in this case.

Personally I have never been able to tell the diference between 44.1, 48 and 96k sample rates except for the amount of space they take up on the Hard Drive. (even if I listen really, really hard with my eyes closed and everything :))
 
I see that you put a great deal of work on this which is great but a couple of things. Firstly alias frequencies are frequencies which are created below Nyquist due to issues like improper filtering of frequencies that are created above Nyquist, not frequencies created due to non-linear processing which exist above.

Hey Tom, thanks for the look. And yes, you're right. It's been a while since I've read about alias freq's. My definition went a little wacky over time. I took it out to avoid confusion/misinformation. Thanks for the correction.
 
First is the idea - if I understand it correctly - is that the higher sample rate can affect the quality of processing and the frequency range in which it it can operate, but that does not automatically mean it's going to inject higher frequencies if the effect does not call for it.

Yeah I expected that myself, as I said in my earlier post. Can't make something out of nothing. I just proved it graphically that there was no higher frequencies/overtones generated through the reverb at a higher sampling rate. This proved true in this scenario. May I say, I have little technical prowess. I think I was more or less doing this as proof of my own assumptions.

I think a better experiment may have been to render the two files with the same reverb plug in, even though it started at 44.1khz, and upsample one to 96khz. Then analyze the difference, if any, between the two, particularly in the high end (where the higher sample rate, might benefit). This may as well result in the same final outcome but might have more merit as an experiment.
 
If you have time I would suggest that you try taking a higher frequency sine wave and running it through a limiter so that it creates something closer to a square wave which will produce harmonics above Nyquist (unless the limiter is removing them as part of its algorithm) and try the same experiment. From there see if frequencies below the original frequency (alias frequencies) are created.

I'm not sure I follow exactly how that experiment would go. I might have to wait for the next time you get to your studio.

I'm stuck in the snow at school, except I have my studio with me. :)
 
I'm not sure I follow exactly how that experiment would go. I might have to wait for the next time you get to your studio.

I'm stuck in the snow at school, except I have my studio with me. :)

Well while not at the studio (and between shovelings) I played around with SoundForge on my laptop. The following shows the results of processing a 44.1k file at the original sample rate versus upsampling the file and then downsampling back to 44.1k:

http://www.masteringhouse.com/upsampling/

Best,
Tom
 
I went through the whole upsampling thing and researched it allot ; came to the same conclusion that it is only advantageous with non-linear processes , and , thats only if your doing some pretty aggressive processing .

Modern converters are pretty dam good , even moderately priced ones are tons better that what you got 15
years ago . Only the cheapo ones on mobo's or boom boxes are totally wretched.
 
I went through the whole upsampling thing and researched it allot ; came to the same conclusion that it is only advantageous with non-linear processes , and , thats only if your doing some pretty aggressive processing .

What exactly is meant when you guys say "non linear processing"? Can you give an example where you can confidently conclude (even w/ aggressive processing) that upsampling has a positive effect.
 
What exactly is meant when you guys say "non linear processing"? Can you give an example where you can confidently conclude (even w/ aggressive processing) that upsampling has a positive effect.

Hi Eric,

A non-linear device is one which essentially distorts the original signal in some way. The output of the device is non-proportional to the input.

Given the graphs on the link I have given previously I can confidently say that this would be a good example of a positive effect of upsampling.

If you notice in the last two figures of the link, sample rate conversion can also cause additional harmonics to be added if the conversion isn't top grade (SoundForge 10 now has a better SRC than the one I had access to on my laptop today, and isn't the SRC that I use for real projects anyway). This is one reason why I recommend to clients that they supply me with mixes that are at the same sample rate as the initial inital tracks if they are mixing entirely in the box.

Given good SRC though, upsampling can definitely be benefit during processing even when the final product is ultimately going to be a lower sample rate just as 24 bit is a benefit even though the final product is 16 bit.
 
Non-linear processes tend to be a reference to dynamics controllers like comressors and limiters. ( MH made his examples were with a limiter ).


Here is the thing ; If you compress a sine wave with a fast attack and release , you can sometimes end up with it being more of a square wave!! The sine wave only has a fundamental and no harmonics . Square waves do have harmonics ( odd ) . So if you look at the pictures MH posted , the " bad : examples had additional "stuff" IE harmonics that were'nt there to begin with around the fundamental freq's whilst the upsampled has much less .


Basically you have to have "aggressive" settings on a comp or limiter to get them to do waveshaping and most times when you are mastering you would'nt have it set so aggressivley ... So all that leaves as an advantage is more samples (slices ) for a limiter with a lookahead function to parse .

The "sooner " you see it gives you a chance to turn it down for an instant ( limit) during the transient ( thats usually where the loudest parts are ). so it's somewhat of an advantage but unfortunatley you are still Waveshaping when you control amplitude in such a short timespace as that( less time than the wave can complete its cycle) .

The whole point is if you don't want inharmonic and/or distortion then you are better off making volume adjustments in a more gentle "volume envelope " fashion ( over a bigger time span ; which is actually more than micro seconds , more like several miliseconds) , which means you won't be causing waveshaping anyways so no upsample should be needed.



To get into why less stuff shows up in the upsampled examples would take more digital theory than I have time for now ( Chicken is ready !!!!!!!!!!!!)


cheers
 
Upsampling also has benefits for processing other than limiting and compression, the UAD precision EQ and Weiss EQ1 for example use upsampling.
 
Back
Top