IPB

Welcome Guest ( Log In | Register )

2 Pages V   1 2 >  
Reply to this topicStart new topic
16 bit, 24 bit and Noise Floor
botface
post Feb 28 2012, 20:13
Post #1





Group: Members
Posts: 354
Joined: 14-January 08
Member No.: 50483



I'm involved in a discussion on a home recording forum. We are not discussing post processing, mixing or mastering just the simple capture of a single acoustic instrument.

I'm suggesting that with a recording environment that yields a noise floor of -60dbfs it makes no difference at whether you capture in 16 bit or 24 bit as you effectively only have 10 bits to work with and so any theoretical benefits of using 24 bit are lost

I'm getting arguments back like :
QUOTE
Specifically 24 bit capture files will still provide the widest range of dynamic and the greatest resolution for capturing anything


QUOTE
With 16 bit you have that many steps and with 24 you have that many more ... which means the more bits you have, the less grainy your resolution is


QUOTE
To give an example: You have a scale of 0 to 100. If you print it on a 10cm piece of metall you barely have enough place to distinguish 1mm distances. If you print it onto a 1m bar, you can have plenty more subdivisions. Of course 100 is still 100 and zero is still zero, but you have many more subdivisions. Same with digital signal. Say you would use a 2 db resolution: then you would have just four different levels to represent the different signal levels which would have to be somehow digitalized onto these four steps.
With 16 bit you have 2^16 possible steps (65536) and with 24 bits 2^24 possible steps (16777216, quite a bit more...). Whether you believe them to be more useful to represent an audio signal I leave to your own ears.I'm not claming that the perceived distance between noise level and maximum level is bigger with 24 bits than with 16 bits. It's just that you have many more subdivisions in between, and that is audible.


QUOTE
If the full theoretically possible dynamical span gets more quantizing levels per dB so must also the same dynamical span of let's say 60dB gain above recorded noise floor. I' can't see where this argument is wrong ... Maybe we're talking about different things?


QUOTE
16 bit recordings have a theoretical dynamic span of 96dB; 24 bit recordings one of 144 dB. If I take the number of possible representations of level with 16 bit, then I have 2^16/96 = 682,7 (rounded) steps per 1dB. WIth a 24 bit recording I have 2^24/144 = 116508,4 (again rounded) steps per 1 dB. So it seems that the dynamical resolution is much finer with 24 bit.


QUOTE
The scale is the same, but with 24 bits we get bigger number of smaller steps than with 16 bit at any part of the scale which allows for better resolution.
You can want to increase the dynamic range by cutting off the noise, and stretching what is left down to the negative infinity level to fill the whole range, and higher resolution of 24 bit will become very handy compared to 16 bit.


I've given examples, I've worked through the arithmetic, I've provided analogies but nobody seems to think the noise floor of the recording environment has any impact on the ability of 24 bit to capture more detail. Can any of you provide any examples or analogies that you've found to work in the past.
Go to the top of the page
+Quote Post
saratoga
post Feb 28 2012, 20:17
Post #2





Group: Members
Posts: 4844
Joined: 2-September 02
Member No.: 3264



You are correct and those people you quoted do not know what they're talking about.
Go to the top of the page
+Quote Post
greynol
post Feb 28 2012, 20:18
Post #3





Group: Super Moderator
Posts: 10000
Joined: 1-April 04
From: San Francisco
Member No.: 13167



As a simple experiment, you can create a pure tone that is quieter than the noise floor and you'll likely find that you can still hear it.


--------------------
YOUR EYES CANNOT HEAR!!!!!!!!!!!
Go to the top of the page
+Quote Post
hlloyge
post Feb 28 2012, 20:21
Post #4





Group: Members
Posts: 695
Joined: 10-January 06
From: Zagreb
Member No.: 27018



Oh, just work in 24 bits so you can have more precise capture. I don't see anything wrong with that. Any postprocessing will be more precise with 24 bit audio.
Theoretically, you can capture 12 bit and still be above room noise. But, why?
Go to the top of the page
+Quote Post
DVDdoug
post Feb 28 2012, 20:26
Post #5





Group: Members
Posts: 2534
Joined: 24-August 07
From: Silicon Valley
Member No.: 46454



Let's look at the "ruler" example -
QUOTE
To give an example: You have a scale of 0 to 100. If you print it on a 10cm piece of metall you barely have enough place to distinguish 1mm distances....


As you mentioned, the problem is noise. Noise is randomness and uncertainty. It's as if those 1mm distances are are jumping-around 10mm. What good does it do to have 1mm worth of resolution when you have less than 1mm of accuracy or repeatability.

People make the same arguement about analog tape or vinyl having "infinite" resolution. But, the noise means that you have less resolution than you get with 16-bits.

On the other hand, all profesional interfaces and most simi-professional interfaces are 24-bits, so you might as well use all 24 bits.

This post has been edited by DVDdoug: Feb 28 2012, 20:29
Go to the top of the page
+Quote Post
greynol
post Feb 28 2012, 20:43
Post #6





Group: Super Moderator
Posts: 10000
Joined: 1-April 04
From: San Francisco
Member No.: 13167



Please, before submitting any more abject analogies, generate white noise clip with an RMS power of -48dB and then superimpose a 1kHz sine wave with an RMS power of -60dB and tell me if you hear the 1k tone.

If you want, I'll gladly provide a sample.

This post has been edited by greynol: Feb 28 2012, 20:50


--------------------
YOUR EYES CANNOT HEAR!!!!!!!!!!!
Go to the top of the page
+Quote Post
KMD
post Feb 28 2012, 20:46
Post #7





Group: Members
Posts: 97
Joined: 21-January 09
From: UK
Member No.: 65825



greynol is right, sounds below the noise floor are audible. Noisy recording processes are additive so using 24 bit to record a noisy enviroment is valid. Anyway the enviroment may be considered to be part of the programe material to be captured.

This post has been edited by KMD: Feb 28 2012, 20:48
Go to the top of the page
+Quote Post
saratoga
post Feb 28 2012, 20:51
Post #8





Group: Members
Posts: 4844
Joined: 2-September 02
Member No.: 3264



QUOTE (greynol @ Feb 28 2012, 14:43) *
Please, before submitting any more analogies, generate white noise clip with an RMS power of -48dB and then superimpose a 1kHz sine wave with an RMS power of -60dB and tell me if you hear the 1k tone.


If I understand correctly, the question is not about hearing below the noise floor, but rather if more bits are somehow "better" independent of the actual dynamic range available. My answer to this is no. Once the quantization error drops significantly below the noise floor of the equipment, its essentially irrelevant.

Apologies if I've misunderstood your point.
Go to the top of the page
+Quote Post
greynol
post Feb 28 2012, 21:22
Post #9





Group: Super Moderator
Posts: 10000
Joined: 1-April 04
From: San Francisco
Member No.: 13167



You're right.

Case in point: create a 1k tone in 24 bits that peaks at -103dB and then convert it to 16 bits with dither.


--------------------
YOUR EYES CANNOT HEAR!!!!!!!!!!!
Go to the top of the page
+Quote Post
mjb2006
post Feb 28 2012, 21:22
Post #10





Group: Members
Posts: 754
Joined: 12-May 06
From: Colorado, USA
Member No.: 30694



Most explanations—even those for technical audiences—of the effect of sample rate and bit depth in sampling generally boil down to the idea of "capturing more detail". I think this is a big part of the problem; everything that people learn about digital audio gets integrated with this basic tenet. You can talk until you're blue in the face about noise, Nyquist, and everything else, but you're still going to run up against this perception: there is a squiggly line which is being segmented along the X and Y axes with a certain degree of precision in order to represent it as a series of numbers, and if you crank up the knobs that control the precision along one axis or the other, you're going to get a better representation of the original squiggly line, so why not do that. When someone's thinking that way, what can you do? You can tell them that if all the music is between 0 and 5, making a knob that goes to 11 instead of 10 is not going to give you "more detail". But they just aren't convinced. So I'm not sure what you can really do. There's still a need for a sampling primer for audiophile myth-busting purposes.
Go to the top of the page
+Quote Post
saratoga
post Feb 28 2012, 21:30
Post #11





Group: Members
Posts: 4844
Joined: 2-September 02
Member No.: 3264



QUOTE (greynol @ Feb 28 2012, 15:22) *
You're right.

Case in point: create a 1k tone in 24 bits that peaks at -103dB and then convert it to 16 bits with dither.


Ah yes I see what you are getting at now.

botface: If you're goal is to show these people how quantization works, perhaps a simple way to do that would be a few example files of pure tones in white noise with a given SNR quantized to different bit depths as greynol suggests.

This post has been edited by saratoga: Feb 28 2012, 21:30
Go to the top of the page
+Quote Post
greynol
post Feb 28 2012, 21:39
Post #12





Group: Super Moderator
Posts: 10000
Joined: 1-April 04
From: San Francisco
Member No.: 13167



I should have led with my second example. wink.gif


--------------------
YOUR EYES CANNOT HEAR!!!!!!!!!!!
Go to the top of the page
+Quote Post
WernerO
post Feb 29 2012, 08:36
Post #13





Group: Members
Posts: 70
Joined: 21-November 06
Member No.: 37858



QUOTE (botface @ Feb 28 2012, 20:13) *
I'm suggesting that with a recording environment that yields a noise floor of -60dbfs it makes no difference at whether you capture in 16 bit or 24 bit as you effectively only have 10 bits to work with


That is probably not quite correct. Assuming that your '-60dBFS' noise level is noise integrated over a meaningful bandwidth (for instance a reading on a level meter) and that you are in a real room with a real system, the spectral distribution of that noise signal is most likely skewed severely to the lower frequencies. That means that once above a few 100Hz you will need more than 10 bits to capture all, simply because the ambient noise there is much lower than the integrated meter reading wants you to believe.


QUOTE (saratoga @ Feb 28 2012, 21:30) *
perhaps a simple way to do that would be a few example files of pure tones in white noise with a given SNR quantized to different bit depths


4 bit demo, including music, here http://www.tnt-audio.com/sorgenti/dither_e.html
Go to the top of the page
+Quote Post
knutinh
post Feb 29 2012, 11:00
Post #14





Group: Members
Posts: 569
Joined: 1-November 06
Member No.: 37047



QUOTE (WernerO @ Feb 29 2012, 09:36) *
That is probably not quite correct. Assuming that your '-60dBFS' noise level is noise integrated over a meaningful bandwidth (for instance a reading on a level meter) and that you are in a real room with a real system, the spectral distribution of that noise signal is most likely skewed severely to the lower frequencies. That means that once above a few 100Hz you will need more than 10 bits to capture all, simply because the ambient noise there is much lower than the integrated meter reading wants you to believe.

A good point. As long as no precise spectral estimate of a)background noise and b)16-bit quantization error is provided and compared, it is difficult to be confident that 24 bits has exacly zero perceptual value.

-k
Go to the top of the page
+Quote Post
Garf
post Feb 29 2012, 15:48
Post #15


Server Admin


Group: Admin
Posts: 4881
Joined: 24-September 01
Member No.: 13



QUOTE (greynol @ Feb 28 2012, 20:43) *
Please, before submitting any more abject analogies, generate white noise clip with an RMS power of -48dB and then superimpose a 1kHz sine wave with an RMS power of -60dB and tell me if you hear the 1k tone.

If you want, I'll gladly provide a sample.


This is slightly surprising to me. Not the effect itself, but the 12dB difference. IIRC, all encoders assume Noise-Masks-Tone thresholds in the neighborhood of 6dB. The white noise will be over the entire spectrum so this shouldn't be an issue of inter vs intra band masking either.

Edit: After looking at this in a spectrogram, it's very obvious. White noise at 48khz with an RMS of -48dB has a spectral floor below -80dB. The tone peaks at -60dB, so obviously it can't be masked. The difference is that the tone concentrates all its RMS power in a single frequency, whereas the noise, while louder in RMS, has to spread that power over all frequencies.

The original post talked about a -60dB noise floor. What does this mean, exactly? If this is a spectral floor, then anything that quantizes from -66dB or better (11 bits) will have no audible advantage. If this is an integrating meter reading, then WernerO is spot on.
Go to the top of the page
+Quote Post
icstm
post Feb 29 2012, 18:56
Post #16





Group: Members
Posts: 121
Joined: 25-January 12
Member No.: 96698



your point about single tones vs white noise is also shown on THIS thread when they are looking CD vs the black stuff.
Go to the top of the page
+Quote Post
Garf
post Mar 1 2012, 08:40
Post #17


Server Admin


Group: Admin
Posts: 4881
Joined: 24-September 01
Member No.: 13



QUOTE (icstm @ Feb 29 2012, 18:56) *
your point about single tones vs white noise is also shown on THIS thread when they are looking CD vs the black stuff.


If you mean that only looking at the RMS level of the signal doesn't tell you much about the noise floor, yes. You don't need to look any further than at DSD to make that very obvious.

The author makes some severe mistakes that void the entire comparison: he's assuming that the dynamic performance of his LP is constant over the input level. This is, AFAIK, absolutely not true due to the physical characteristics. This can be seen clearly at the end of the linked page, and to the authors credit, he admits this in his conclusion (but somehow claims its an advantage - quite the unbiased comparison there). The entire comparison would also be broken if there is so much as a smooth lowpass filter present on the LP, which, given the admitted harmonics distortions, would be quite likely anyway. Lastly, he investigates silence on a digital CD and concludes it has an RMS power of -103dB. I have no words for this. The RMS power of a silent CD signal is minus infinity. Zip, nada.

He's also trying to draw some conclusions from the peak sample level, which is just...sigh.

This all a bit tangential to our discussion. The original posters statement of "a recording environment that yields a noise floor of -60dbfs" just doesn't convey enough information to know whether recording at more than 10 bits makes sense or not. If this noise floor is caused by a 50Hz transformer hum at -60dB, and the spectral floor at all other frequencies is significantly lower, then recording at 24 bits may certainly provide advantages, especially if there are multiple mastering steps involved. If this noise floor exists because of broad-spectrum ambient noises, he might as well use a tape recorder.
Go to the top of the page
+Quote Post
botface
post Mar 1 2012, 10:13
Post #18





Group: Members
Posts: 354
Joined: 14-January 08
Member No.: 50483



QUOTE (Garf @ Mar 1 2012, 07:40) *
The original posters statement of "a recording environment that yields a noise floor of -60dbfs" just doesn't convey enough information to know whether recording at more than 10 bits makes sense or not. If this noise floor is caused by a 50Hz transformer hum at -60dB, and the spectral floor at all other frequencies is significantly lower, then recording at 24 bits may certainly provide advantages, especially if there are multiple mastering steps involved. If this noise floor exists because of broad-spectrum ambient noises, he might as well use a tape recorder.

Garf,
I talking about the ambient room noise, mic self noise, preamp noise etc so it just general broadband noise.

@everybody else,
Thanks for the responses
Go to the top of the page
+Quote Post
2Bdecided
post Mar 1 2012, 10:51
Post #19


ReplayGain developer


Group: Developer
Posts: 5057
Joined: 5-November 01
From: Yorkshire, UK
Member No.: 409



Try lossyWAV with --bitdist to see how many bits it thinks you need to keep.

Cheers,
David.
Go to the top of the page
+Quote Post
Garf
post Mar 1 2012, 11:19
Post #20


Server Admin


Group: Admin
Posts: 4881
Joined: 24-September 01
Member No.: 13



QUOTE (2Bdecided @ Mar 1 2012, 10:51) *
Try lossyWAV with --bitdist to see how many bits it thinks you need to keep.


Preferably with options that don't assume the quantizer is noise-shaping, though smile.gif He also can't control the quantization on a block-by-block basis, so he'd have to use the worst-case.
Go to the top of the page
+Quote Post
Ethan Winer
post Mar 1 2012, 19:59
Post #21





Group: Members
Posts: 248
Joined: 12-May 09
From: New Milford, CT
Member No.: 69730



QUOTE (greynol @ Feb 28 2012, 14:18) *
As a simple experiment, you can create a pure tone that is quieter than the noise floor and you'll likely find that you can still hear it.

Sure, but not if the signal is 96 - 60 = 36 dB below the noise, using the OP's example of room noise which is typical. Of course, all noise is not uniformly broadband, but the basic principle is valid. You can hear maybe 10 dB below the noise, but certainly not 36 dB unless the noise is sculpted to remove frequencies present in the source.

--Ethan


--------------------
I believe in Truth, Justice, and the Scientific Method
Go to the top of the page
+Quote Post
Arnold B. Kruege...
post Mar 2 2012, 01:24
Post #22





Group: Members
Posts: 3641
Joined: 29-October 08
From: USA, 48236
Member No.: 61311



QUOTE (knutinh @ Feb 29 2012, 05:00) *
QUOTE (WernerO @ Feb 29 2012, 09:36) *
That is probably not quite correct. Assuming that your '-60dBFS' noise level is noise integrated over a meaningful bandwidth (for instance a reading on a level meter) and that you are in a real room with a real system, the spectral distribution of that noise signal is most likely skewed severely to the lower frequencies. That means that once above a few 100Hz you will need more than 10 bits to capture all, simply because the ambient noise there is much lower than the integrated meter reading wants you to believe.

A good point. As long as no precise spectral estimate of a)background noise and b)16-bit quantization error is provided and compared, it is difficult to be confident that 24 bits has exacly zero perceptual value.


The general point to be made is that a dB for dB masking of a lower amplitude noise floor only happens when both noise floors have identical spectral shaping. However, just because the two noise floors have different spectral shapes does not mean that all bets are off.

The right thing to do is to subtract the two on a point-for-point, octave-for-octave or whatever relevant means of comparison you have available , and then apply an appropriate audibility weighting curve to the difference. Since the levels of noise floors are pretty low, the roll off of high and low frequencies will usually be pretty extreme. This tends to minimize the original differences in the shapes of the spectral content of the two noise floors.

Having measured a ton of noise floors over the years, i've seen just about every rule of thumb to fall apart. Probably the canonical noise floor of a room will be red or brown noise-shaped, but a heavy dose of turbulence in the air flow of the HVAC system can wash a lot of that out. Some microphones have noise floors that are shaped more like white noise, while other are shaped more like pink noise. Complementary equalization curves (e.g. RIAA) can distort the noise floors of electronics between the pre-emphais and de-emphasis networks. Finally, a 60 dB noise floor is almost a whopping 40 dB away from a 96 dB noise floor and that tends to wash our a lot of minor differences.

Also, while the noise floor of real world 16 bit systems is usually within a few dB of 96 dB, very few real world 24 bit systems come within 20 dB of 24 bits. Usually, going to 24 bits from 16 is only good for a 10-20 dB advantage and sometimes it is as little as 5 dB.

Go to the top of the page
+Quote Post
Glenn Gundlach
post Mar 2 2012, 07:36
Post #23





Group: Members
Posts: 364
Joined: 19-April 08
From: LA
Member No.: 52914



QUOTE (greynol @ Feb 28 2012, 11:18) *
As a simple experiment, you can create a pure tone that is quieter than the noise floor and you'll likely find that you can still hear it.


I have a Nakamichi 580M cassette deck with a factory claimed S/N ratio of 65dB. I also have a Sony test CD that a friend wanted copied to cassette. The disc includes 1KHz tones 0, -10 down to -90dB in 10 dB steps. It was easiest to simply record the cassette in one pass even though many of the tones were below -65. And yes, the machine was properly set up and the reference level was correct. I played it back as a joke and was very surprised to hear ALL the tones start and stop including -90dB. No ABX was needed as everyone could hear it clearly. 1/2 lsb dither produces similar results in digital but I haven't tested at 25dB below the noise floor though it might be an interesting experiment.

Go to the top of the page
+Quote Post
Garf
post Mar 2 2012, 12:07
Post #24


Server Admin


Group: Admin
Posts: 4881
Joined: 24-September 01
Member No.: 13



QUOTE (Glenn Gundlach @ Mar 2 2012, 07:36) *
I have a Nakamichi 580M cassette deck with a factory claimed S/N ratio of 65dB...but I haven't tested at 25dB below the noise floor though it might be an interesting experiment.


A S/N ratio of 65dB does not imply a flat noise floor -65dBfs. I'd expect a cassette deck to have a better S/N ratio at lower frequencies than at high ones. This is why Dolby B/C works.

As far as I know, you can't hear tones more than about 6dB below the noise floor.
Go to the top of the page
+Quote Post
Arnold B. Kruege...
post Mar 2 2012, 14:02
Post #25





Group: Members
Posts: 3641
Joined: 29-October 08
From: USA, 48236
Member No.: 61311



This is another one of those situations where your mileage varies all over the map depending on the details. One common-sense requirement is that the tone be audible when there is no noise.

The spectral content of the noise is a very important parameter. In audio systems composed of just electronics, the spectrum of the noise floor can commonly range from pink to white to the sort of peaked-up just below Nyquist spectrum that results from using perceptually-shaped dither. The actual listening level used in the evaluation is also very important.

It is common for people to judge these things with the signal level boosted by 40 dB or more, as compared to a normal gain setting based on comfortable listening to normal music. This asymmetry is very popular with people who think they have just invented the perfect new dither that will solve all of the problems of digital. This throws things so out of whack in comparison with actual practical use that it isn't even a bad joke. ;-)

This post has been edited by db1989: Mar 2 2012, 14:07
Reason for edit: removing unnecessary full quote of above post
Go to the top of the page
+Quote Post

2 Pages V   1 2 >
Reply to this topicStart new topic
1 User(s) are reading this topic (1 Guests and 0 Anonymous Users)
0 Members:

 



RSS Lo-Fi Version Time is now: 22nd July 2014 - 13:04