How to hear samplerate conversion degradation?

[Phazma]

So, it sometimes happens that I have to deliver projects in other samplerates than I worked in or I need to load samples from different libraries with different samplerates into my projects.
It is all pretty much a mess but I love that Reaper takes and delivers seamlessly what it is asked to.

However I know that there should not be unnecessary SR conversion happening. And when doing polarity-inversion-null-tests with for example 44.1 and 48.0 renders of the same project (resample mode always Extreme HQ 768pt HQ Sinc) the high frequencies don't really null eachother out.

But when AB-ing the renders and trying to hear differences by ear I am not able. As if it was the same file.

So I am wondering is the conversion degradation negligible from a non-mathematical point of view and may I proceed to work on whatever SR I see fit and with a mix of files and just as a final step render what is asked to deliver or do I need to put extra caution into this.

When listening for differences between different SR conversions, what should I focus on to detect differences and reliably decide for myself if they do or don't do my work justice?

Also if I have for example to deliver a project at 48KHz and most sounds I see fit to use are recorded at 44.1 or 96, should I still work at 48KHz and let Reaper convert the single sounds or work at the SR that most sounds are at and convert to 48KHz when rendering?

Or is it best to convert sounds even before they are imported?
In past I had to mostly work at 44.1 and got most sounds according to that but now more and more it happens that I work at 48Khz. Should I batch-convert my sounds so they are all 48KHz when importing?

Lots of questions, hope someone can clear up the best practises when it comes to working with various SR and how to spot possible degradation if conversion is unavoidable.

Edit:
Also, is it better to have "Use project sample rate for mixing and FX/Synth processing" ticked or unticked if conversion during rendering is necessary?

[DVDdoug]

If you don't have a choice you don't have a choice so it doesn't really matter. Just try to eliminate unnecessary multiple conversions. Even then you might have to make 10 conversions before there is noticeable degradation. And of course, your listeners can't A/B.

But when AB-ing the renders and trying to hear differences by ear I am not able. As if it was the same file.

That's normal. A lot of people think they hear a difference but in a scientific, blind, level-matched ABX Test those "obvious" differences usually disappear...

the high frequencies don't really null each other out.

When you listen to the null file can you hear anything without cranking-up the gain?


If it nulls, that proves there's no difference and your done!


But if it doesn't null it doesn't necessarily mean anything. The sound of the difference isn't the same as the difference in the sound. For example, if you add a few milliseconds of silence to the beginning to make a short delay that doesn't change the sound at all but subtraction makes a very-loud comb-filtered difference file. And although you may recognize comb filtering as the result of delay it doesn't sound like delay.

[Phazma]

Thanks for the answer, sounds reassuring.
So if I have to work on a project that needs to be delivered in 48KHz and majority of the audio files I use are at 44.1KHz my best bet is to work at 44.1KHz and just render to 48KHz and nobody will notice?

Thing is, I can indeed hear the high frequencies pretty clearly. Here is a comparison of the spectrum of one of the renders against the difference of both renders.





As this is pretty far from a clear null I was suspecting that I should be able to hear some difference, yet I can't. And I also checked at maximum zoom, the renders do perfectly align in phase and the waveforms look pretty much identical to my eyes.

I was just concerned that me having a messy workflow with different audiofiles at different samplerates in the same project, perhaps even rendered to yet another samplerate could become one of those things that might go unnoticed but add up and expose themselves as a problematic down the road. Also because I recall having read in some thread that Reaper's SR conversion isn't that trustworthy and that some people use dedicated software for SR conversion but that sounds like a lot of hassle to me.

However I never come even close to 10 conversions (2-3 at max, unless Reaper converts destructively when switching back and forth between project samplerates). So if that means I do not need to worry than that's great news for me. Just gotta love Reaper for absolutely not being picky.

[cassembler]

It's probable that no one will notice and no one will care if they do notice, my $0.02. If you're doing 90-track orchestra work for Disney's Frozen 3, maybe they'll start asking questions...

See this, you may very surprised. (Unfortunately, they don't have a sample rate test, but I'm sure you can find one.) I was, frankly, shocked at pretty much every test I tried (particularly bit depth):

https://www.audiocheck.net/blindtests_index.php

EDIT: Question - are you using any random-based effects (reverb?) That itself would explain the null difference. Upsampling from 44.1 to 48 should literally only add blank stuff between ~20.5k and ~24k. (If you upsample _after_ render, that's the result I'd expect.)

[Phazma]

Originally Posted by cassembler

EDIT: Question - are you using any random-based effects (reverb?) That itself would explain the null difference. Upsampling from 44.1 to 48 should literally only add blank stuff between ~20.5k and ~24k. (If you upsample _after_ render, that's the result I'd expect.)

I can't remember for sure about the project whose spectrum I posted but I think I did convert with the batch converter after render.
Anyway I have done another test by converting another track from 44.1 to 48 and in that case the audibile difference when inverting polarity of one of the renders is indeed only some crunchy noise on the transients (it extends down the spectrum but the high frequencies are louder than the lows). Maybe there is some difference also due to the audio material, the previous one was a more orchestral work and the second one more EDM kind of.
However I will not obsess about this and just keep working at the sample rate at which most used files are and later convert to another SR if necessary.

[planetnine]

Phazma , could you keep your images smaller please..?

>

[cassembler]

Originally Posted by Phazma

I can't remember for sure about the project whose spectrum I posted but I think I did convert with the batch converter after render.
Anyway I have done another test by converting another track from 44.1 to 48 and in that case the audibile difference when inverting polarity of one of the renders is indeed only some crunchy noise on the transients (it extends down the spectrum but the high frequencies are louder than the lows). Maybe there is some difference also due to the audio material, the previous one was a more orchestral work and the second one more EDM kind of.
However I will not obsess about this and just keep working at the sample rate at which most used files are and later convert to another SR if necessary.

Out of curiosity, I attempted this test, and I see what you're saying. Similar results. I suspect it's due to my testing methodology:

1) Take an existing 32/44.1 audio file and convert to 32/48 using Wavelab
2) Invert the polarity on the 32/48 file
3) Add each file to a separate track (this itself triggers yet another sample rate conversion, in my case back down to 44.1)
4) Export the file

I didn't see anything over about -68dBfs, and what's there is gritty, mostly 20k+, BUT I was surprised at how much was below 20k. Not enough to lose any sleep over, but also non-zero. (I seriously doubt any human could AB/blind the difference.)

Academically, I'm wondering now if it's even possible to null-test a sample rate conversion using only a single conversion pass, without needing a second conversion to again match the two streams...