It is what it is. A little bit of drama and conflicts even keep it more fresh. I don't mind being labeled as lazy or annoying or whatever as long as it's interesting and helpful
I still have a lot of experiments to do using all the information in this topic. But I have some intermediate results

If you want to export 48 kHz project to 44.1 kHz, don't just set 44.1 in export settings like me in the first post of this topic. Completely reload the audio engine of your DAW for mastering project to be in 44.1 kHz and then also choose 44.1 kHz in the export settings
Just compare the same track:
1. 24 bit/48kHz in project settings -> exported to 24 bit/48kHz WAV = -0.11 db peak and -0.02 dbtp
2. 24 bit/44.1kHz in project settings -> exported to 24 bit/44.1kHz WAV = -0.11 db peak and +0.12 dbtp
3. 24 bit/48kHz in project settings -> exported to 24 bit/44.1kHz WAV = 0.00 db peak and +0.74 dbtp (true peak)

If both project and export settings are 44.1kHz (variant 2), we get the ordinary peak at the same level as if we exported in 48 kHz. And a slightly higher true peak value (for whoever that matters)

However, the last time limiter and all mastering chain worked in 48 kHz, so inside the DAW there were no peaks. But then it got converted to 44.1 khZ and we got ordinary peak increase in the resulting file. Still not a big difference, but if I hadn't set the ceiling to -0.1 in limiter settings, I would have gotten +0.1 ordinary peak (that's unacceptable anywhere). And I believe that exact thing might happen if you feed the perfectly fine 48 kHz file to distributor without testing it this way. So I highly recommend performing the test number 3 if you have a ton of loud projects in 48 kHz like me

Next we move to bit rates testing. All of those had 48 kHz in project settings, but different sample and bit rates in export settings to emulate streaming conversions:

48kHz/24bit and 48kHz/16bit had exactly the same peak values (both ordinary and true) = -0.11 db peak and -0.02 dbtp
44.1kHz/24bit and 44.1kHz/16bit also had exactly the same peak values between themselves = 0.00 db peak and +0.74 dbtp (true peak)

So as we can see, in my case bit depth DOES NOT affect peak/true peak value upon conversion. It's only reduced sample rate that gives me unexpected increase of peaks

I may sound like Captain Obvious for somebody, but these are revelations to me. Strange that I didn't get the response like "you noob, bit depth doesn't affect your peak deviations, don't even bother" before. That one sentence would be far more helpful than deleted "100 paragraphs of valuable advice"

 

I asked a pro mastering engineer whether they deliver in 16 or 24 bit format and this was their reply:

 

This is very valuable. Thank you!
My "fear of missing out" gets confirmed by this. Especially regrading Apple Music and Amazon. Never heard of Qobuz before though. Looks like a smaller thing for audiophiles, but still.

96k is out of question for me - it's going to be just pointless upconvert (unless I switch to recording audio in 96k in future).
But that quote might be the reason for some people here to think about switching to 24bit/48k

I also watched this video by Dan Worrall and he explained why higher sample rates can be even worse in some cases (due to increased intermodulation and stuff). But 48k seems to be a great spot. (25:45 timecode)

 

In case you're interested in the physics behind it, there's an article from Xiph–the developers of Vorbis and FLAC (they know their stuff...)–on why everything above 44.1/48 kHz and 16 bit is pointless or even harmful for music distribution:
https://people.xiph.org/~xiphmont/demo/neil-young.html

 

Last edited: Jul 30, 2023

yikes

 

Yeah i doubt most guys deliver at 96k tbh. I know for sure there are engineers that deliver in 44k. I think the guy i emailed is particularly OCD with it. But i guess that's the point, everyone's different, there's no real right or wrong answer, and it wont really affect the soul of the song in any way, and that really illustrates what i think is THE answer to your original question: there's no right or wrong answer. Just release it

 

I'll take a look. And it works for me without web archive. But I actually wanted to come back to your earlier message.

I thought this part was worth discussing:

But then everything went into big drama show with our beloved keyboard warriors

 

Last edited: Jul 30, 2023

Oh yeah. You are the dude that told me to "Go educate yourself". Then edited the message to just smiley face. And then to just a dot. Eventually it was removed. What was that? Nerves?
You and Haas are either the same person from different accounts or attending the same masterclass called "Spotify is Devil. Stop before it's too late"
I wish I've typed "Apple Music" or "Streaming services" instead of Spotify. Maybe then you all could discuss the actual topic instead of my horrible personality. Or even better - not bother us with your presence

 

For sure. I'm in no hurry until the video is ready, so out of curiosity I'll still mess around with that Apple and Nugen utilities. And see how many more Spotify crusaders we can enrage with this topic

 

Today's discoveries

SRC from 48 to 44.1 kHz normalizes everything to 0 db. At least in Bitwig. That's why I got 0 db in every 44.1 file I exported in the tests before. And all the 44.1k FLACs from the platforms will be at 0 db, even if they got clipped and weren't actually at 0 db before SRC

If I set my limiter ceiling to - 1db in 48 khz project, I get the following:
-1.01 db peak in 48k exported file
-0.57 db peak in 44k exported file

For -2db ceiling:
-2.01 db peak in 48k exported file
-1.57 db peak in 44k exported file

So as you can see, I have the increase of 0.44 db in ordinary peaks after 48 to 44.1 SRC every time (no matter the ceiling). And if the limiter's ceiling is less than -0.44, they just get clipped by Bitwig's normalization to 0 db.
The exact same thing can happen with distributors and streaming platforms if you don't pay attention and blindly feed them 48k

0.44db peak level difference is a lot. And what if some platforms have worse SRCs than Bitwig? You can end up with even more peak clipping before your file even gets encoded to AAC, OGG, MP3 or whatever. So you may be surprised by the results afterwards

 

SRC induces a steep/brickwall low-pass filter to minimize the effects of Shannon-Nyquist DSP theory drawbacks' which is aliasing. In theory only though as it's hard for SRC engines to create such a "perfect" digital filter without artefacts such as ringing at the cutoff point.

When you do oversampling which is upsampling (creation of "points" for the reconstructed digital waveform) + interpolation (those points gets connected to create the analog waveform), you increase the max frequency of the signal and the number of samples (44100 samples per second compared to 48000 etc). But when you go past the maximum frequency of the signal (22k50 for 44100 Hz) you get "foldback" of the signal, hence aliasing.

When you downsample, the peak level changes due to those digital points being cut out (you're substracting samples to fit the new sample rate). The phase and magnitude of the new downsampled signal changes hence the increase of peaks.
That's why on Izotope RX SRC' module as an example, you have the option to toggle "post-limiter conversion" which allows to watch for those nasty peaks which add up to the signal.

 

Last edited: Aug 1, 2023

I emailed another pro mastering engineer asking if he delivers files as 16bit or 24bit and this was his answer:

I generally trust this guy more than the other guy I quoted, at least as far as technique is concerned, he has better credits a too, but I think he might be mistaken about Spotify using MP3.
Regardless, I think this further illustrates the point that everybody is different and there's no right or wrong answer.
It does nothing to create any more certainty for us lol. But there you go

 

If I understand correctly, filter ringing at the cutoff point won't be an audible problem (if filter introduces it at 22k50 and not somewhere below). But we can probably consider it as a kind of aliasing too if it is able to increase peak level around that freqency

In my case oversampling is used only in Maximizer. And I think it takes care of the aliasing after oversampling and then getting back to project sample rate of 48k
But I believe what you say here is exactly what's happening during SRC after exporting a file. Frequencies in the range of 22.5 to 24k get folded back because they are present in 48k file, but can't be present in 44k file. So seems like aliasing from 22.5 to 24k is giving me that +0.44 db peak increase in the resulting file after 48 to 44 SRC

If the problem lies only in that inaudible area of 22.5 to 24k, then theoretically I can just high-shelf it minus a few dBs and see what happens after SRC in that case

 

Thank you again! I totally understand the position of this mastering engineer. But it's interesting how the first one you asked cares about HiFi options and this one consciously buries those
So he basically leaves no choice for his clients to use the enhanced potential of Amazon Music Unlimited, Tidal HiFi Plus, Qobuz and Apple Music Lossless unless they bring that up

In the ideal world it would be great to upload separate versions for each scenario: 16bit/44k for standard stuff and 24bit/48k for audiophiles. But distributors won't let that happen, as far as I understand.

So you need to make a choice:
1. Lose almost half a dB of loudness to be safe on 48 to 44 SRC but benefit from being present on HiFi plans
2. Master in 44k and ditch the HiFi plans

I'm personally looking for the third option - optimizing the 48K file to benefit from both options (not lose any loudness and appear in HiFi subscription plans). And seems like I'm getting closer to the solution

I also found another opinion on Logic forum reagrding it. Attaching the screenshot

 

You could but you would loose some of the definition or "air" you added to the high frequencies while mastering. That's why you should theorically only care about those fold-backed frequencies when applying SRC or not using oversampling over various treatments.

Any audio treatment inducing high-frequency phase or gain change will trigger aliasing effects no matter what. As you said it's in the "inaudible range" and you're correct but that's why many engineers work in the native sample-rate provided by the client's mixes (let's say 48 kHz) then downsample it with various SRC engines according to their preferences (Izotope RX, Weiss Saracon, etc).
You work in "full-definition" of the signal then take care of the digital audio drawbacks when downsampling.

Also think of high frequencies as "higher number of CPS : Cycles Per Second" (instead of Hz for this exemple), those are hard to reproduce when you're approaching the maximum frequency of the signal. Hence the implementation of this "ideal" LP steep filter to make sure no cycle gets added back to the signal when you substract samples.

 

Last edited: Aug 1, 2023

Understood, thank you. But you know what's crazy?
I'm comparing 44k vs 48k in mastering project settings right now. And having the same master chain in 44k results in more open highs then 48. The rest of the spectrum is identical
I only have eq+limiter on vocal bus and mastering eq+limiter on master bus in the final project. So only those 4 plugins running in 44k
give such a difference