Skip to main content

Notice

Please note that most of the software linked on this forum is likely to be safe to use. If you are unsure, feel free to ask in the relevant topics, or send a private message to an administrator or moderator. To help curb the problems of false positives, or in the event that you do find actual malware, you can contribute through the article linked here.
Topic: 'Normalization' of PCM audio - subjectively benign? (Read 140749 times) previous topic - next topic
0 Members and 2 Guests are viewing this topic.

'Normalization' of PCM audio - subjectively benign?

Hi All,

A thread back in March discussed whether normalization of WAV files is 'lossy', and I remember that no-one thought to point out that operative word is actually 'destructive'.

Digital processing of PCM audio such as normalization, compression (or expansion), equalisation, etc' is 'destructive', not 'lossy', although both mean 'irreversable'. It has to be said, this seems as much a matter of semantics or even philosophy as much as exact engineering terminology.

Anyway, this is of some interest to me, I'm using a 24-bit ADC to record from vinyl (and taking it pretty seriously as an archiving project - I get one shot at some discs, so I want to do it right), and as it happens there is  insufficient gain on the line from the phono-amp I'm using to it to get close to 0dB with some discs, many peak at -6, -8, even -12 dB.

However, even -20 dB or more in 24 bits is greater than 16-bits of resolution. A 24-bit recording which peaks at, say, -16dB, can be 'normalized' to 0dB, and will *still* have information below digital silence in 16-bits, and ideally require dither on conversion.

Obviously I could use a preampifier in addition to the phono-stage to get around the level issues, but I'd rather keep to the 'minimal' signal path I'm using now.

What I'd like to ask people here is; in their experience, is normalization completely 'benign', sonically? Are the algoritms used in different applications much the same, or are some better than others?

R.

'Normalization' of PCM audio - subjectively benign?

Reply #1
Theres nothing worth recording anywhere near 120dB below peak on a record, so you're not losing anything by normalizing.  There would only be a loss if the dynamic range of the album plus the wasted room at the top were greater then your ADC was capable of recording.

Personally I wouldn't use normalization since it doesn't really get you anything replaygain doesn't do.  So I'd just replaygain the albums and then apply it during playback (if desired).

'Normalization' of PCM audio - subjectively benign?

Reply #2
At least peak normalisation should be more or less the same in every application because it's a rather simple process (I guess). But it's always possible that there are differences. And keep in mind that it is irreversible.

You can find a short description if you search for normalization in the scientific R&D forums here and most likely a lot about it on wiki or in the web out there.

But I would also recommend you using replay gain if possible since you don't need to touch the audio data and it works with perceived loudness, not peak amplitude or rms (well it uses rms afaik, but it's just a part of it). http://wiki.hydrogenaudio.org/index.php?title=Replaygain

'Normalization' of PCM audio - subjectively benign?

Reply #3
The thing is, I don't want to end up with CDR's that peak at -12dB or lower.

By normalizing at 24-bit resolution, some info is effectively 'pulled' up into the usable dynamic range of CD-audio.

To reiterate; the question is, has anyone noticed the effects of the rounding errors in normalization, and are all nornalization algorithms created equal?

A few years back I tried 'remastering' some late 80's CD's that were typically at -6dB or lower (there seemed to be a lot of paranoia in those days about hitting 0dB) , and I thought (NB) they didn't sound as good as the original files.

Rounding errors in 24-bits will be less of concern, presumably

'Normalization' of PCM audio - subjectively benign?

Reply #4
Yes, it is completely "benign,"  at least within reasonable limits. Working in floating point gives you the best margin. The amplification will not in any way damage the audio, the only serious consideration is the signal to noise ratio. If your recorded signal is low enough, the noise floor of your system will become an audible part of the finished product.

This is not too likely to be a significant factor with a decent 24 bit soundcard, but at some point, if the amplification of your phono stage is low enough, you will gain quality by using a decent mixer or line level preamp to increase the signal in the analogue domain.

I believe this is a fairly simple calculation as digital transforms go, although I have no idea but that there may be some program(s) that makes a (relative) mess of it.

'Normalization' of PCM audio - subjectively benign?

Reply #5
I use replay gain for my own listening (with Foobar, all my music is on my PC), but for serious lights down and a spliff listening, it's kernel-streaming and no DSP of any kind.

In the context of straight LP to CD transfers, I don't want to use more digital processing than absolutely necessary, and as you might have gathered I'm even a little suspicious of simply adding gain.

'Normalization' of PCM audio - subjectively benign?

Reply #6
The thing is, I don't want to end up with CDR's that peak at -12dB or lower.

By normalizing at 24-bit resolution, some info is effectively 'pulled' up into the usable dynamic range of CD-audio.


Well something is pulled up into the dynamic range of the CD, but its not info.  Probably just hiss, since the dynamic range of the CD is so much greater then the record (even including 10 or 20 dB of waste).

Edit:  I am dumb.

'Normalization' of PCM audio - subjectively benign?

Reply #7
I guess that's the conundrum - add a pre-amp or add gain digitally.

Probably you're right, and I should really get the analogue side optimized. The only problem is that simple, decent and affordable preamps are few and far between (no market for the little suckers, I guess)- I have thought of building a simple opamp based one myself.

'Normalization' of PCM audio - subjectively benign?

Reply #8
I guess that's the conundrum - add a pre-amp or add gain digitally.

Probably you're right, and I should really get the analogue side optimized. The only problem is that simple, decent and affordable preamps are few and far between (no market for the little suckers, I guess)- I have thought of building a simple opamp based one myself.


I think you don't understand me.  I'm saying it makes absolutely no difference since the limit is the source not the hardware

'Normalization' of PCM audio - subjectively benign?

Reply #9
Quote
I have thought of building a simple opamp based one myself.
good option but take care of old opamp like 741 that have round -40dB noise floor and with the power supply to your pre ampl or you'll get more noise than quality. if you don't choose the right components,better is amplify or normalize in 32bit and after that back to 16bit.

'Normalization' of PCM audio - subjectively benign?

Reply #10
Well something is pulled up into the dynamic range of the CD, but its not info.  Probably just hiss, since the dynamic range of the CD is so much greater then the record (even including 10 or 20 dB of waste).


Well, the whole point of dithering 16-bit PCM is that sounds like tape-hiss, for example, (say at -70 odd dB) can sound rather gritty and 'unnatural' without it.

It's been argued that analogue noise (whether from tape or circuitry), which is essentially 'brownian', shouldn't be any more intrusive than other 'ambience noise'. I'd rather have that tape hiss quantized with as many bits as possible.

It's generally understood very low-level signals (such as ambience cues) aren't reproduced well by CD-audio near the bottom of it's dynamic range - simply not enough bits..

'Normalization' of PCM audio - subjectively benign?

Reply #11

Well something is pulled up into the dynamic range of the CD, but its not info.  Probably just hiss, since the dynamic range of the CD is so much greater then the record (even including 10 or 20 dB of waste).


Well, the whole point of dithering 16-bit PCM is that sounds like tape-hiss, for example, (say at -70 odd dB) can sound rather gritty and 'unnatural' without it.

It's been argued that analogue noise (whether from tape or circuitry), which is essentially 'brownian', shouldn't be any more intrusive than other 'ambience noise'. I'd rather have that tape hiss quantized with as many bits as possible.



I don't think you're understanding this.  The source media you have has maybe 60 or 70dB of dynamic range.  If you're lucky.  The CD you're going to has roughly 100dB, perhaps more with noiseshaping.  Thats 30-40dB of headroom.  Applying dither here isn't going to help you since the bottom 5-7 bit are already randomly distributed and thus contain no actual information. 

It's generally understood very low-level signals (such as ambience cues) aren't reproduced well by CD-audio near the bottom of it's dynamic range - simply not enough bits..


You're not even close to the bottom.

'Normalization' of PCM audio - subjectively benign?

Reply #12
Quote
In the context of straight LP to CD transfers, I don't want to use more digital processing than absolutely necessary, and as you might have gathered I'm even a little suspicious of simply adding gain.

Having done more than 400 LP transfers I submit that reasonable digital processing is no detriment to audio quality. Of course it depends on what you are looking for, but it is easy to produce a product through editing that is better than what came off the LP.

Reading posts about DEA, it seems as though more than a few people are so concerned about the bits they end up with -- far beyond anything that might make an audible difference -- that one might believe they are building their eternal abode in paradise instead of a mobile music collection. Well, whatever you might think about that attitude, it is very silly to try to apply it to recordings you make from an LP. Record the same disk twice and the totally unavoidable differences between the two recordings will be far more significant than any "damage" you could possibly do by normalizing.

Since peaks too near 0dB can produce waveforms as much as 8dB above 0dB (although that extreme is pretty unlikely in music), some DACs can clip on files that have no clipping. I recommend normalizing to about 97% instead of going to maximum.

'Normalization' of PCM audio - subjectively benign?

Reply #13
I don't think you're understanding this.  The source media you have has maybe 60 or 70dB of dynamic range.  If you're lucky.  The CD you're going to has roughly 100dB, perhaps more with noiseshaping.  Thats 30-40dB of headroom.  Applying dither here isn't going to help you since the bottom 5-7 bit are already randomly distributed and thus contain no actual information.


Analogue sources have no '0dB' reference to allow direct comparison of 'dynamic range' or 'signal to noise' (which two terms are used, wrongly, interchangably).

However, distortion in the analogue 'domain' increases with level, in the digital domain *exactly* the opposite is the case.

Now, LP. You say 60-70 of 'dynamic range', when in fact you meant 'signal to noise', and that would be about right.

However, since information can be stored, retrieved and heard *well* below that noise floor, the dynamic range is actually considerably greater.

Properly mastered, and played back on commensurately good equipment, LP actually has a dynamic range approaching 90dB, and all of this reproduced with distortion of no more than a few %, and this subjectively 'benign' distortion at that.

'Normalization' of PCM audio - subjectively benign?

Reply #14
Having done more than 400 LP transfers I submit that reasonable digital processing is no detriment to audio quality. Of course it depends on what you are looking for, but it is easy to produce a product through editing that is better than what came off the LP.

Reading posts about DEA, it seems as though more than a few people are so concerned about the bits they end up with -- far beyond anything that might make an audible difference -- that one might believe they are building their eternal abode in paradise instead of a mobile music collection. Well, whatever you might think about that attitude, it is very silly to try to apply it to recordings you make from an LP. Record the same disk twice and the totally unavoidable differences between the two recordings will be far more significant than any "damage" you could possibly do by normalizing.

Since peaks too near 0dB can produce waveforms as much as 8dB above 0dB (although that extreme is pretty unlikely in music), some DACs can clip on files that have no clipping. I recommend normalizing to about 97% instead of going to maximum.


Amongst others, I will be transferring discs my Dad will be sending over from AUstralia (all classical), many of which date back to the 50's and 60's. It's a one-time thing in the case of these LPs, and I would like to do as good job as I can.

Doing a 'good job' always invloves attention to detail, as I'm sure you're aware.

Anyway, on the subject in hand - normailzation - I mentioned that I tried applying this to some rips of overly 'quiet' 80's CD's, and I wasn't impressed with the results - I went back to turning the volume up.

That was of course 16 bit - I'm hoping that the process is more benign at 24 bits.

I agree that you can 'improve' on the LP itself, in that discreet pops ans clicks can be competely masked, and it would be lost opportunity not do do so. I'm less sanguine about removal of other kinds of noise, though.

'Normalization' of PCM audio - subjectively benign?

Reply #15
Quote
Analogue sources have no '0dB' reference to allow direct comparison of 'dynamic range' or 'signal to noise' (which two terms are used, wrongly, interchangably).
This is wrong, actually. Vinyl has always had a "reference" level of 0db, referring to a maximum stylus velocity, which, depending on who you ask, is either 5cm/s or 7cm/s peak. From what I understand, cutting levels are calibrated by how they would reproduce a 1khz sine wave, which would have been at an amplitude of 0db on your tape, relative to that level. (EDIT: Just like tape, this 0db reference can be exceeded, with perhaps greater distortion.)

You tend not to ever think about that 0db reference level if all you're recording is 12" LPs. 12" EPs and singles are often cut at much higher levels (+3 or +6db). Test records can quite commonly hit +15db, and a few hit +18db as a torture test. But if you want to actually get an accurate signal-to-noise result with a turntable system, it's probably a good idea to compare against 0db.

I'm pretty much in the same boat as you are with recording LPs at low levels. Most of my recordings trade at around -15 to -20dbFS peak - which is both worse than yours and perfectly reasonable. It would be foolhardy to boost my gain by 15db, because I do have several tracks that legitimately hit -8 to -10db.

To boot, I have a full-discrete RIAA preamp, and it sucks. I think it's worse than the preamp in my receiver. When I have extra time and pennies I'll put together an OPA637-based amplifier. Nowadays it is quite possible to design an opamp preamp with lower noise and distortion than many discrete designs.

Quote
However, since information can be stored, retrieved and heard *well* below that noise floor, the dynamic range is actually considerably greater.
I made the exact same argument a few months ago and I got shot down to hell for it. As I believe somebody (Garf? Woodinville?) told me, to paraphrase, according to modern psychoacoustic theory, at any frequency with a given level of noise, you're only really going to hear a signal that's at most 3-6db underneath that noise. What this means is that even though a signal "seems" to be buried in the noise and is yet still audible - a signal that peaks at -80dbFS being audible behind pink noise at -70dbFS, for instance - those numbers don't mean anything unless they are broken out by frequency. And when you do that you will find that the signal is usually going to be well above the noise level.

Quote
Properly mastered, and played back on commensurately good equipment, LP actually has a dynamic range approaching 90dB, and all of this reproduced with distortion of no more than a few %, and this subjectively 'benign' distortion at that.
Again, the dynamic range figures are highly frequency dependent. At 10hz I would estimate that you would be extremely lucky to get 40db out. If you test only at 1khz, and go from +15db all the way down to whatever is measurable, then I guess you could get 90db. But that number is going to be wildly different at different frequencies. Personally, when I've looked at spectrum plots of my sound card noise versus blank vinyl grooves, the vinyl/preamp noise dwarfs the internal soundcard noise by a very wide margin at all frequencies.

'Normalization' of PCM audio - subjectively benign?

Reply #16
Unless you used a very poor program, the only undesirable thing you could achieve by normalizing is making the existing noise more noticeable. Turning up the volume in the analogue domain will do exactly the same thing. You should proceed in whatever way makes you happy, but you are laboring under a fantasy if you think there is a difference.

With 16 bit files it is possible to either simply truncate the calculation results or dither them. Dithering significantly reduces distortion, and thus is to be preferred, at least in a conceptual sense. With LPs  however (or cassettes), the background noise level is so high relative to the quantization errors from truncating that there isn't any contest. You will never hear the difference (unless you do a sufficiently large number of transforms; normalization is one transform).

It is easy enough to demonstrate this. You can record into a 32 bit floating point format and do all operations there. The quantization errors will all be very far below the theoretical limits of circuit noise floors -- thus there is no way to possibly hear them. Then you resample to 16 bit (with or without dither is almost certain to be audibly undifferentiated). You can then take the initial recording and immediately convert to 16 bit, then duplicate the operations there. You won't be able to differentiate the two in a blind AB test. You can add to this testing by using 24 bit integer files in any manner you want to manipulate them. The results will be audibly identically.

Clicks and pops are not "masked" by any process I'm familiar with. They are removed, either completely or to a large enough extent that they can no longer be distinguished. In the majority of cases there is nothing left to suggest they were ever there.

Removing many other noises is a process with so many variables that many people seem to never get it right. That doesn't mean that it can not be done very well indeed. While it isn't possible to do blind AB tests, since any audio needing noise reduction will sound different after processing, it is possible to do close tolerance AB comparisons. After years of doing so I am quite convinced that my most general results are that there is no significant change in the desired audio, the noise level is just reduced. In cases of very poor condition LP, the best results may indeed make some audible changes to the signal, but the results are thus better than the beginning condition, no contest.


***

In any practical terms, the dynamic range of LP is the difference between the highest amplitude signal and the lowest that can be differentiate from the intrinsic disk surface noise. While some people have speculated that greater signal depths can exist, actual measurable dynamic range is about 55 to 60 dB; in fact in many cases the recording is compressed to that neighborhood in the mastering process. With some fairly heavy duty noise reduction it might be extended to 70 dB in a transfer from LP. Where have you heard/read of anyone measuring anything approaching 90dB?

Every analogue device (any used in audio, anyway) has a linear range. These are routinely graphed in the specification sheets. A properly designed circuit keeps operation in this range. Distortion does not increase with signal level unless the circuit is over driven.

'Normalization' of PCM audio - subjectively benign?

Reply #17
I guess there is a difference in subject matter. I'm writing about dynamic range of any given recording. Axon, and I presume Rockfan, are writing about the possible dynamic range of the medium. NO?

'Normalization' of PCM audio - subjectively benign?

Reply #18
Maybe I can put it this way...

If the noise floor of your pre-amp + ADC is well below the noise floor of the most "silent" parts on a given disc for all audible frequencies, then you are transferring the full dynamic range of that recording (as reproduced by your turntable) into the digital domain.

To determine the noise floor of your pre-amp + ADC, try recording the output of your pre-amp when the stylus is not on the record. You might be surprised - I'm betting that there's already noise well above the 16th bit before you even play the LP! (Though maybe not at all frequencies).


In your application, normalisation is perfectly benign. It's a trivial mathematical operation. That doesn't mean all program do it without error, but they should!

When converting from 24 to 16-bits, you should dither.

I agree with other posters that there may be no audible benefit from working at 24-bits, or from dithering. However, given that you already have the hardware and software to do both, they don't cost anything, and can't hurt.

Cheers,
David.

'Normalization' of PCM audio - subjectively benign?

Reply #19
...While some people have speculated that greater signal depths can exist, actual measurable dynamic range [of an LP] is about 55 to 60 dB; in fact in many cases the recording is compressed to that neighborhood in the mastering process. With some fairly heavy duty noise reduction it might be extended to 70 dB in a transfer from LP. Where have you heard/read of anyone measuring anything approaching 90dB?

There have been numerous examples of frequency spectrum graphs showing apparently very low noise levels from LPs once you get above about 500Hz. The only link I can find to hand is this article on the Audioholics website.

However, the real noise floor is not as low as it appears in these graphs. Rather than repeat the details here, if anyone is interested in why I believe that the conclusions drawn are in error, see this section of my LP-to-CD tips page.

'Normalization' of PCM audio - subjectively benign?

Reply #20
I'm glad you've critiqued that, Clive.

That's the difference between a simple RMS measurement, and one of any number of possible frequency dependent measurements.


Simple RMS is clear enough to define and measure, but can be misleading.

A strictly defined "spectrum level" is also clear (though surprisingly small for most broadband signals). If something (e.g. noise) is spectrally flat, take the RMS level, go from dB to linear, divide by the bandwidth of the noise, and convert back to dB. That's the spectrum level.

When you FFT something (as Cool Edit does), you're not getting the correct spectrum level. The results depend on the window function, the window length, and the reference level.

If you use a shorter window, then you spread information in the frequency domain - this reduces the apparent (peak) level of sine waves. Conversely, if you use a longer window, then you average energy over time while examining smaller frequency bands - this reduces the apparent level of broadband / noise-like signals relative to sine waves.

If you know what you are doing, you can prove almost anything!

And that's even before you compare RMS measurements with a poor estimate of spectrum level - as you point out, two completely different things!!!


The relevant comparisons are those which compare like with like, and use percentually relevant parameters (time resolution, frequency resolution) for whatever is being measured.

Cheers,
David.

'Normalization' of PCM audio - subjectively benign?

Reply #21
What David said. Once you move to spectral measurements you have to move away from RMS, peak or A-weighted noise numbers entirely, and work off of your own measurements, for the most part. They're not at all standardized.

Oh, and I'll throw another curveball at the original question: If you only amplify your signal in ~6db increments - ie, if you repeatedly double your signal until going any further will clip - no quantization error will occur. It will be a 100% lossless process. Of course, you can only get away with this when you can specify your signal amplification as an integral ratio instead of as dB of gain, because it's impossible to specify an exact factor of 2 increase in db.

'Normalization' of PCM audio - subjectively benign?

Reply #22
off topic:
this thread is "bookmarked"!

2Bdecided,Axon,cliveb,AndyH-ha,RockFan,Hollunder,Mike Giacomelli,
you are now in my "white book"(hall of fame).
you all posted magnific explanations in few words,is the HA best thread in my taste.


'Normalization' of PCM audio - subjectively benign?

Reply #23
That's very useful info, Clive, thanks for sharing your experience. And BTW, I agree about the 'demise' of Cooledit.

Re. the adequacy of 16-bit for recording LPs; essentially the question more broadly is whether 16 bits is adequate for music, period, and I would agree that it is.

However, if any DSP is going to be carried out, IMO 16 bits is really NOT adequate. This kinds of relates to my question about simple normalization.

An example I vividly remember was applying normalization to an old copy of "School's Out", which was absurdly quiet. At first I was quite happy with the results, but something wasn't right (to be suitably vague), and I ended up binning it. Possibly flaws in the infamously dodgy A/D conversion often used in in the early days of CD had been emphasized.

You mention the need for 'hard measurement', and while no-one would argue with that, one has to recognize the fact that the 'art' (!) of music reproduction is not a fait-accomplis. Music is easily the most complex and cerebral use we make of our auditory system, and some 'parameters' have yet to be nailed down (IMO). Even the best audio engineers/designers in the world will listen to their designs before deciding whether they're any good at playing music.

PS >> my own 'standards' for transferring vinyl are similar to yours - I'm not a "audiophile extremist" but as I've repeatedly said, I want to do it right - I regard it as 'archiving.' I'm using a completely overhauled, DC-driven Townshend Rock MkII (with a cassette-tape belt, believe it or not), heavily-modded/rewired/screened  Rega arm and Garrott-tipped Decca MkIV, a Gram Amp 2SE and the aforementioned stand-alone 24-bit ADC, the latter both battery-powered.

'Normalization' of PCM audio - subjectively benign?

Reply #24
However if any DSP is going to be carried out, IMO 16 bits is really NOT adequate. This kinds of relates to my question about simple normalization.

In the context of LP transfer, I am not persuaded of the need for working at greater than 16 bits. When you consider that the resolution of an LP is about 10 or 11 bits (12 on a good day with a following wind), then you'd need to do a heck of a lot of DSP for the rounding errors to accumulate sufficiently to bring the 16 bit quantisation noise up above the vinyl noise floor.

And the point here is that, IMHO, LP transfers don't need much DSP at all. The vast majority of the editing I do concerns the removal of impulse noise, which involves edits to small isolated sections of waveform. The changes made to the waveform at those locations vastly swamps any changes to the quantisation noise that may result.

The only "global DSP" I do to LP recordings is normalisation (nearly always), some modest EQ (very rarely), and broadband noise reduction (sometimes, and only in moderation). Apart from normalisation, these other operations again make a much bigger change to the audible nature of the music than the minor change in quantisation noise they cause (which I still maintain will remain beneath the vinyl noise floor).