Normalization methods for LP digitization

[danadam]

Morover, doing it 96-->48 instead of 88.2-->44.1 allows you to apply downsampling not too steeply, letting the 20Hz - 20kHz range completely untouched.

Anything I should be worried about, assuming I'm using a competent resampler (i.e. sox)?

 

[Lttlwing16]

In Audacity, the setting is "Effect | Amplify". When you apply it, it uses whatever dither you selected in settings. After applying it, you will see the waveform get bigger or smaller, depending on whether you used a + or - value. If you export to FLAC (or WAV) after doing this, it saves the modified version.

From the Audacity documentation here it seems as though peak normalization and amplify function the same way at their core function, but have different methods about getting there. In your example, Amplify just sets the peak at -1.0 dB, calculates the dB difference from the current peak, then adds or subtracts that to all the data. Amplify lacks DC offset, and always raises/lowers amplitude equally on both channels of a stereo track, thus preserving the original channel balance. Normalization gives one the option for DC offset and allows optional independent adjustments per channel.

I didn't see anything in the documentation about the use of the dither to carryout these functions, but I find the language about down-sampling in the Audacity manual somewhat confusing. Audacity Dither Manual

 

[solderdude]

capture at a lower peak, around -6.0 dBFS, then use peak normalization to bring each side of the LP up to around -1.0 to -0.5 dBFS.

Record peaks around -6dBFS at 24/96 (192 is overkill), remove clicks/ticks/noise at least the ticks that peak above the music, normalize to -1dB, split.

 

[Lttlwing16]

Why not use the same capture and normalize it two ways?

Where does alsamixer have VU meter? Or did you mean arecord?

Sorry forgot to include that important bit of info -- methods 1 and 2 are from the same master capture. Method 3 was recorded in arecord, and it's VU meter peaked at 95% (not alsamixer, thx!). Fixed both those above.

Oh, and assuming it shows peak value and it is exactly 95%, this should translate to about -0.45 dBFS. Maybe that would explain the slight difference between 3 and the other two.

For whatever reason alsamixer will only allow me adjust capture gain at intervals of 2, so that was as close I could get without clipping.

That's somewhat strange. Do you have a way the measure the LUFS loudness of the peak normalized file?

Not currently, although it seems there is a LUFS meter available in the Arch Linux repository. I'll give it a shot tonight. If you have an Audacity setup, give it a go and see if you get the same result, i.e. matching dB RMS, but one noticeable louder.

I'm not sure I see the connection.

Recording at the record's peak without processing would be capturing the mastering engineers target LUFS, and sound profile, or as close to it as my hardware and software can allow. I can confirm this if I can get a functional LUFS meter and measure the LUFS of the track as close to peak capture as possible. If the target is indeed close to -17 LUFS, then this would explain the sound similarities between the EBUR128 normalized capture and the peak un-processed capture.

Any difference in fidelity between those files should be only from the fact that they are three separate captures, e.g., in one capture a click or pop happened while in the other it did not, or was quieter. If there actually is any fidelity loss from applying +/- 6 dB gain in a 24 bit file, assuming no clipping happened, I would start looking for something that is broken . Unless the ADC is so bad that its noise floor is at vinyl record level and that 6 dB makes a difference.
But that's just me

My fault, two captures : ) .. Could be, but I can state I also had the same conclusion on a separate track from a completely different LP, with the exact same test.

The Realtek ALC262 data sheet on pg 72 shows the ADC with -82 dBFS THD+N and S/N of 90 dBFSA.

 

[rdenney]

I wonder if you are overthinking this. I don't try to characterize the music by averages--that seems to me like it would require a lot of consideration to avoid either clipping or leaving unused headroom. I use Vinyl Studio set to peaks around -6 dB or lower and normalize both side of an album together such that the peak is at -0.2 dB. If it's not as loud as modern processed music, I know where the volume knob is.

I don't want to expand or compress the dynamics. At. All.

The dynamic range of the phono input in my preamp is around 95 dB, and my Benchmark ADC is much better than that. The full dynamic range of an LP is maybe 70 dB (which is already a fantasy), so I can record 10 dB down on the peak, normalize to -0.2 dB, and the system noise will still be far below the vinyl roar. In any case, on playback, my needledrops sound exactly like the LP being played back directly, even played loudly, if I don't load it down with a bunch of noise processing crap. If I heard something harsh, it would be because some intermediate amplifier is clipping, I suspect.

My equipment chain is: Audio-Technica AT440Mla cartridge to Adcom GFP-565 phono preamp input to record bus (tape loop output) to Benchmark ADC1 USB to Vinyl Studio in computer. Playback is computer to Musical Fidelity V90 DAC to the tape monitor input to the preamp. I compare by switching back and forth between phono input and tape monitor input. The preamp would be the weak link if it wasn't 25 dB better than the cartridge.

Rick "who wants to hear the LP as is, even when it's bad" Denney

 

[MRC01]

From the Audacity documentation here it seems as though peak normalization and amplify function the same way at their core function, but have different methods about getting there. In your example, Amplify just sets the peak at -1.0 dB, calculates the dB difference from the current peak, then adds or subtracts that to all the data. Amplify lacks DC offset, and always raises/lowers amplitude equally on both channels of a stereo track, thus preserving the original channel balance. Normalization gives one the option for DC offset and allows optional independent adjustments per channel.

I didn't see anything in the documentation about the use of the dither to carryout these functions, but I find the language about down-sampling in the Audacity manual somewhat confusing. Audacity Dither Manual

DC offset removal is simply a high pass filter. With vinyl, this kind of filter is also useful for reducing subsonic rumble/resonance from the LP recording. You can do it yourself, just apply something like corner frequency 20 Hz, slope -6 dB / octave. That will get you both.

That: amplitude shift and high-pass filter, is the only processing I did with my LPs. I preferred not to apply dynamic compression, noise or pop/click reduction, etc.

 

[sarumbear]

Hey folks,

I've been exploring digitizing my LP's so I can play them on my upstairs stereo where a turntable setup isn't an option. I wanted to open a discussion about methods to properly bring in the analog data and have it at a reasonable playback volume. In particular normalization methods.

From the research I've done so far, while most agree a high bit depth at capture is best, it seems people are split on capture method otherwise. Here are some of the methods I've seen proposed/used:

• capture at a lower peak, around -6.0 dBFS, then use peak normalization to bring each side of the LP up to around -1.0 to -0.5 dBFS.
• capture at as close to -1.0 to -0.5 dbFS, no normalization
• capture at a lower peak (-6.0 dBFS), no normalization, adjust output volume with playback device (volume control)
• capture at a low peak then use REPLAYGAIN tag to adjust volume

The lower capture volume seems important to avoid clipping/distortion and remove any added THD+N from the soundcard itself, although the noise floor of the LP is most likely much higher than that of the soundcard.

The problem with normalization, and peak normalization in particular, is the presence of artificial peaks from record noise (i.e. clicks and pops). While one can meticulously groom the waveforms to remove them, some are still liable to be present and can then affect the normalization process. Additionally, this adds an additional layer of processing to the digital capture. I assume this is why some don't care for it.

However, I came across this article which explains the three different methods for normalization: peak, RMS, and EBU-R128. The latter two methods were interesting, and I had come across RMS normalization in Pro Tools, but never understood what it did.

My question is then, why isn't anyone discussing normalizing to an appropriate LUFS, instead of the peak dBFS? For instance, I own a digital copy of Khruangbin's Mordechai as well as the LP. I could import the FLAC files and check the LUFS level on each track, then using the Loudness Normalization tool in Audacity (which utilizes the EBU-R128 volume detection method) to bring the LP rip to the same perceived LUFS?

This is how archivists at Abbey Road do.

1- Record at -10dBFS using a 20-bit resolving ADC at 24-bit 48kHz.

2- Post process the clicks using a custom plug-in that detects the clicks and reduce their level by 15-20dB. The software works by looking at the decay of the peak and if it’s longer than the rise leaves it alone. That’s because in music decay is slower then rise due to reverberation.

3- Normalise the file to -0.5dBFS

Edit: Corrected DAC to ADC

 

[Lttlwing16]

Interesting: I came across this from the AES Technical Document AESTD1008.1.21-8 recommendations on Loudness page 10:

 

[abdo123]

RMS is a kind-of "average" so it's an indication of "loudness" but it's just "simple" calculation and it doesn't take frequency content into account like LUFS or ReplayGain.

RMS is basically what the amplitude of the entire signal would be if it was DC current.

 

[sarumbear]

Interesting: I came across this from the AES Technical Document AESTD1008.1.21-8 recommendations on Loudness page 10:

View attachment 156094

I can’t see the relevance of that paper. Besides, why do you want to apply compression while archiving?

 

[daftcombo]

Anything I should be worried about, assuming I'm using a competent resampler (i.e. sox)?

Probably nothing audible. But doing simulations in iZotope, using the integrated resampler, which is very good, and with 88.2-->44.1, it is a tradeoff between touching the upper side of the 20Hz - 20kHz range or adding noise afterwards.

 

[solderdude]

Record at -10dBFS using a 20-bit resolving DAC at 24-bit 48kHz.

They are probably using an ADC... it is very hard to record an analog signal using a DAC

 

[sarumbear]

They are probably using an ADC... it is very hard to record an analog signal using a DAC

Corrected the post. Thank you

 

[sarumbear]

Interestingly AES recommends -16 LUFS for track normalization, which method 2 was slightly below. Obviously album normalization would be preferred here to preserve the album dynamics from song to song. Album Dynamic is recommended to -14 LUFS. Read more about AES recommendations here.

You are confusing broadcasting/streaming with origination. Archive is origination. You are simply changing the media. You do not want to change the master with compression.

 

[Lttlwing16]

I can’t see the relevance of that paper. Besides, why do you want to apply compression while archiving?

Sorry @sarumbear I didn't see anything about adding compression to the archival process from the post you quoted.

The article is relevant as it lays out guidelines for proper loudness adjustments to music.

 

[sarumbear]

Sorry @sarumbear I didn't see anything about adding compression to the archival process from the post you quoted.

The article is relevant as it lays out guidelines for proper loudness adjustments to music.

I am afraid you are mistaken. Loudness control is a process. You never do any processing while archiving. Removing clicks or corrective equalisation are accepted as they are corrections to the media's deficiencies, in this case clicks and turntable limitations.

Read the title of the paper you referenced, I didn't reference any paper.

 

[Lttlwing16]

I am afraid you are mistaken. Loudness control is a process. You never do any processing while archiving. Removing clicks or corrective equalisation are accepted as they are corrections to the media's deficiencies, in this case clicks and turntable limitations.

Read the title of the paper you referenced.

You stated Abbey Roads archival process includes normalization to -0.5dB.

Perhaps I'm misunderstanding, but normalization would be considered processing.

 

[sarumbear]

You stated Abbey Roads archival process includes normalization to -0.5dB.

Perhaps I'm misunderstanding, but normalization would be considered processing.

Normalisation of the whole album is volume control. Loudness matching track to track is process. Allow me to eloborate.

Streaming and broadcast has limited dynamic range capacity to work on suboptimum listening environments. A classical piece has a crest factor around 17dB. Quite passages are 17dB lower than the high passages. Imagine that level of difference in a car. You have to constantly adjust the volume to hear every passage at the same perceived level. At a domestic audio system noise is much less and such level difference works OK. Piano parts play at a low level, forte parts at a high level. We feel the difference in level as the composer wanted. On streaming and broadcast those different parts must be loudness matched. If you do the same while archiving you are effectively changing the master. Naturally, this doesn't apply to modern pop music but as we are talking about vinyl digitising I expect modern pop is not the criteria here.

It is important to differentiate between broadcast, recording and mastering.

 

[rdenney]

You stated Abbey Roads archival process includes normalization to -0.5dB.

Perhaps I'm misunderstanding, but normalization would be considered processing.

Normalization doesn’t change the dynamics or the coloration. It’s just a straight digital amplification. Setting the loudness level repositions the peaks with respect to the average, as I understand it, and that is processing—it changes the dynamics.

The distinction is the same in photography. When scanning film for archival storage, we apply corrections to fix the errors in the scanning device, and that’s based on a calibration process. We also adjust levels so that the brightest highlights are just short of absolute white. And we correct flaws like dust that show up in the scan. That maintains the separation of tones for archival storage, but does not apply any interpretation.

We might then interpret the archival file artistically, to make the image look like we want it to look.

Finally, we target the image to carry that interpretation to the display device or print (whatever that might be) as faithfully to the interpretation as possible. That’s like compressing music for playback in a car, which to my thinking should be done by thd playback system in the car, but certainly not to the master used for home playback.

Trying to mix those steps leads to a lot of tail-chasing and also leads to losing touch with what the original was like.

The Beeb not normalizing is reasonable, but it isn’t necessary—normalizing doesn’t cause the original to be changed. We could, of course, set the levels so that the peaks are at -0.5 dB (or whatever) when making the needledrop, like we used to do when recording cassettes, but it requires knowing where the loudest peak is in the music. Clipping digitally is a disaster requiring starting over. So, recording at a somewhat lower level and then normalizing is a lot more convenient. It’s just a matter of when in the process we “set the levels”.

Rick “don’t clip and let the average fall where it may” Denney

 

[MRC01]

Essentially, we're talking about 2 alternatives.

1. After recording, shift the amplitude of the digital recording (the entire LP, not track by track) so it peaks at -1 dB. (or -0.5). Safe the shifted digital file.

2. After recording, scan the digital recording to find the average & peak levels, then create metadata for replay gain, so it can shift the recording level on playback.

Method 1 is guaranteed transparent but is more work. Method 2 is less work, but may or may not be transparent, depending on the implementation of replay gain used during playback.