Using normalise.

[Mr D]

If you're worried about maintaining the most resolution, why not (instead of normalizing your stems on export) just export 32 bit files? That way you have the benefit of keeping your relative balances intact.

TBH i'm always skeptical about exporting 32 bit, as even the difference between 24 and 16 bit is hard enough to hear, let alone 32 bit.

A few people have mentioned valid uses of normalization, but i don't think it has much place in production because each sound in a mix needs a different level, so choosing some arbitrary starting level seems like an unnecessary extra step.

[jamcat]

I conjecture that it's not destructive,
it does not take any fidelity away
(...what information is lost?)

What is lost is the original information. A "destructive edit" is any edit that permanently changes the original source audio, as opposed to real-time processing. If you do something to the audio and then overwrite the original file with the altered file, that is a destructive edit, regardless of the subjective quality of the change.

It's a good idea not to do that.

[BONES]

You have to watch out for ISPs (inter-sample peaks) though, as these are not detected by standard normalising algorithms. This is where the actual samples are still technically below 0dB, but there is actually a hidden peak between them that will clip the reconstruction filter. This is because samples are actually "connected" by curves rather than straight lines. Because everything is sine waves.

You do understand that's not a real thing, right? Your audio editor just draws those curves, they don't actually exist in the audio file. They are just a convenient visual representation that's easier to take in than a bunch of dots.

Normalizing allows high fidelity by maintaining bit-depth

Not really. You can't create bit depth so if your signal is very quiet, normalising it will add a lot of interpolated data that may or may not be 100% accurate.

...and of course, the last thing that happens to a master is a normalize

Second last. The last thing will be to down-convert your working bit depth to your final output bit depth.

Honestly I don't know why anyone would ask a question here. This is one of the most basic, simple questions and almost nobody has given a useful answer. The display of ignorance around such a basic subject is breath-taking and more than a little bit scary.

The only useful answer is that you should do it if you are going to downsample to a lower bit depth, as it will preserve more of the original signal. So if you have a 24 or 32 bit file you are going to convert to 16 bit, it's always a good idea to normalise first. Beyond that, it probably won't matter one way or the other, except in very specific, extreme circumstances. I tend to do it a lot, mostly as it makes it easier to see the waveform, which saves me having to preview the audio as often. It also makes it easy to swap one sample for another without having to worry about one being louder than the other. It saves me a bit of time when I am working but it doesn't affect the quality of my rendered files at all.

[Mr D]

Honestly I don't know why anyone would ask a question here. This is one of the most basic, simple questions and almost nobody has given a useful answer. The display of ignorance around such a basic subject is breath-taking and more than a little bit scary.

The only useful answer is that you should do it if you are going to downsample to a lower bit depth, as it will preserve more of the original signal. So if you have a 24 or 32 bit file you are going to convert to 16 bit, it's always a good idea to normalise first.

Why would you bit-reduce stems you exported or files you imported? Makes no sense. That's why no-one mentioned it. Straw-man criticism.

[jamcat]

You have to watch out for ISPs (inter-sample peaks) though, as these are not detected by standard normalising algorithms. This is where the actual samples are still technically below 0dB, but there is actually a hidden peak between them that will clip the reconstruction filter. This is because samples are actually "connected" by curves rather than straight lines. Because everything is sine waves.

You do understand that's not a real thing, right? Your audio editor just draws those curves, they don't actually exist in the audio file. They are just a convenient visual representation that's easier to take in than a bunch of dots.

No, that’s not correct. Samples are taken at regular intervals, and there is a good chance that it won’t happen at exactly the local maxima. Instead, the 2 closest samples to the peak may be on the left and right slopes.

Do you really think the peak is just truncated? Because it is not. The reconstruction filter, also known as the “smoothing filter” does just what the name implies. It creates a smooth continuous analog signal. So when the sound is reproduced, it is not a series of discrete dots, or even stepped stairs, but instead a smooth continuous signal that follows sine curves.

This is why different samples rates will sound identical so long as the frequencies being reproduced are below Nyquist. If the D/A conversion didn’t result in a smooth curve, there would be a difference in sound quality due to the larger ‘gaps’ at lower sample rates.

And not all audio editors draw the curves. Most that I’ve seen just okay connect the dots and create trapezoidal waveforms. Ones like Acon Digital Acoustica that show curves do so to represent what the reconstructed sound wave will be doing. It’s not done for convenience, and actually takes more work to code.

[whyterabbyt]

You have to watch out for ISPs (inter-sample peaks) though, as these are not detected by standard normalising algorithms. This is where the actual samples are still technically below 0dB, but there is actually a hidden peak between them that will clip the reconstruction filter. This is because samples are actually "connected" by curves rather than straight lines. Because everything is sine waves.

You do understand that's not a real thing, right? Your audio editor just draws those curves, they don't actually exist in the audio file. They are just a convenient visual representation that's easier to take in than a bunch of dots.

He's not talking about the files, or the visual representation of the files in the DAW. He's talking about the resultant audio signal after digital to analog conversion. That's why he mentioned the reconstruction filter.

https://www.earlevel.com/main/2023/01/1 ... ple-peaks/
https://techblog.izotope.com/2015/08/24 ... detection/

[talby]

Normalization for me is a recommended step where afterwards automations shall be taken advantage of.

Using track templates, I normalize all sources placed on the track to the always same value, for instance to -6 dB peak value. Like this my plugins already placed on the track after the template placed them there, will receive the signal by comparable input levels. Consequently, most preadjusted plugin settings do not have to be readjusted, if working on similar projects for which the templates have been created for.
When recording with the same mics and audio-interface settings, but the acoustic instrument may have been positioned slightly different to the mics, then after normalization of the recordings even preadjusted noise reduction parameters usually fit right away the expectations. Also, if you know how to sidechain your multiband EQ and multiband compressor, then even frequency and dynamic corrections will benefit from the by automation obtained win of speed for getting your mix worked out. And normalized recordings as a starting point do not hinder you to afterwards still become creative in the mix.

[BONES]

Shakers at full scale? No thanks!

If you want the best audio quality, you need to take up all the available bit depth. I use normalise before downsampling. That way I can save all my samples at 16 bit and they'll all be CD quality without taking up masses of drive space.

[osiris]

I only use Normalize when I have a very low volume track. Got a piano loop that was nice but recorded very low. Normalizing really made it shine. But as said don't normalize everything. You can achieve much by just cloning the audio. (Watch for phase issues)

[sin night]

I avoid normalization, because it’s a destructive editing when performed on fixed point files (it may be performed at 32-64 bit floating point, but then it’s saved in the file).

On the final mix, once I “process it for loudness” (I can’t call that “mastering”), I normalize, of course (or, rather, I set the limiter for a consistent level).

If I’m recording a bunch of drum samples for later use, I’ll probably normalize them (or adjust the gain by ear) in order to avoid big changes in volume when auditioning / using them in my projects… the workflow wins in this case, so I apply the destructive editing (note: I don’t record/save in floting point format).

When I work inside a daw, I’d rather use the object handles, the channel strip gain or a gain plugin and keep those adjustments non destructive / floating point.

As a rule of thumb, I’d avoid any destructive editing… But when I need it, I just use it. But it has to be a purpose. A single destructive editing won’t matter much (at least at 24 bit… and depending on the actual recording), but it’s important to avoid accumulating destructive editings when they are not needed.


Anyway, normalization is just gain, there’s no quality gain…

[sin night]

Speaking of audio quality…

The dynamic of a signal is often less than the avaible dynamic of a converter. When I use my hardware synths, it’s not unusual to see some movement (noise!) on the meters at, let’s say, -90dB / -80dB while the gain is already appropriate to get a loud level and there are still some dB of headroom before clipping.
So, the signal often has less dynamic range than the converter (even when considering just 16 bits).

Of course keeping the recording levels too low is not a good idea, because it would further lower the dynamic range on the way in (say my recording peaks at -60dB….), but also raising the level (via analog or digital gain) won’t have any benefit to the dynamic range (once the gain is enough to record the whole dynamic range of the source). If the signal has 70dB of dynamic range, it won’t really matter if the recording peaks at 0dB or -10dB, both recordings should be within the 96dB allowed by 16bit PCM…

So, my approach is to try to keep the noise floor as low as possible (get the full dynamic range kf the sound) while reaching a reasonably high level but still keeping some dB of headroom for further processing.

Trimming down the gain after recording is not an issue for me, but I try to record with a level which is both appropriate for the plugins (because of the analog emulations…) and allows me to keep the faders towards unity (because that’s where they are most precise/sensible).

With tha said, if it’s something meant to go inside a sampler, then I may normalize / apply gain in order to fit my workflow.


Jusr my 2 cents. And maybe other approches are as much effective comparwd to mine…