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: Dynamic range in high fidelity systems (Read 52201 times) previous topic - next topic
0 Members and 1 Guest are viewing this topic.

Dynamic range in high fidelity systems

Reply #50 –
Quote from: Sycraft on
...So suppose you said "11 bits should be enough," and you make a sample like that and it sounds fine. Ok, now what happens when we want the sound level to be 25% of the original, -12dBFS? Well that signal only has 9 bits of resolution, the 2 MSB are not used. So if we have a recording with transients just 10 or so dB above average, which is not at all uncommon for properly mastered classical or the like, we have quite a low resolution for most of the data.

Noise is another issue.
No, in a correctly dithered system, noise is the only issue wrt bit depth. This "quite a low resolution for most of the data" is meaningless other than in the sense that "most of the data sits close to the noise floor" - and even then, it's irrelevant unless the noise floor is audible.

Claims that much of the musical information is sitting in, say, the last 4 bits "therefore it must sound really awful" aren't born out by reality.

Audiophiles like to believe it - but it's totally unsubstantiated. It's quite strange they believe it, because they also love SACD where all the music is sitting in a single bit, completely below the noise floor, and without correct dither! Yes, most of the junk (non-linear distortion and noise) is moved to higher frequencies, but still...

Quote
As such when storing in digital is behooves us to use a resolution such that it can contain the entire dynamic range we want to reproduce, and maintain sufficient resolution at any level within that range.

That's what 24-bit gets us over all.
For almost any and every commercially released recording, that's what 16-bit gets us over all!

Cheers,
David.

Dynamic range in high fidelity systems

Reply #52 –
Quote from: Yee on
Yes - resolution vs. dynamic range. It's two different concepts altogether, not just confusion over terms. Thanks for clarifying.


????????????

Shannon joined dynamic range and resolution with an equation.

Quote
In terms of resolution - I am thinking maybe we can prepared down sampled tracks of music with different resolution - 15bits, 14bits, 13bits... - and ABX with the original and find the higher resolution where we can hear a different. But then the resolution of the ear is likely to be lower than the speaker so we ended up measuring our hearing instead.


This has been done, and the usual answer is that with normal music, decreases in resolution start becoming noticable around 12-13 bits. One problem is that recordded music usually has no more than 14 bits of resolution.

Quote
As for dynamic range - I am still mystified how it could be measured on any reproduction medium.


Record the standard test signal for resolution on that medium from a source with far greater dynamic range than the medium, and then play the test signal back and analyze what you get.

Quote
Take speaker for example - how does one determine where the noise floor is?


The most obvious spurious responses from a loudspeaker are usually due to nonlinear distoriton.

Quote
The bits unit is obviously not relevant here, as even we can somehow establish that the dynamic range is 120dB - there is no way to tell whether that is 10 bits (2048 discrete steps) or 11 bits (4096 discrete steps) in between.


120 dB iis about the same as 20 bit resolution.
 

Dynamic range in high fidelity systems

Reply #53 –
Quote from: Arnold B. Krueger on
Quote from: Yee on
Yes - resolution vs. dynamic range. It's two different concepts altogether, not just confusion over terms. Thanks for clarifying.


????????????

Shannon joined dynamic range and resolution with an equation.


On itself, without reference to the signal source, digital encoding is very well defined theoretically. But if the source has noise floor higher than 1 bits and/or actual dynamic range less than 16bits, then Shanon equation would become quite meaningless.

The same arguments applies in playing back the digital signals. The playback equipments may have all sorts of limitations that makes whatever noise floor/dynamic range of the 16bits/44.1kHz recording quite meaningless.

I was drawn into this discussion because the Harbeth designer made a statement that Harbeth speakers has an 'resolution' equivalent of 11/12 bits bit-depth. On further reflection - Shanon equations obvious cannot apply, as a speaker is not a digital system.

But the statement was based on some empirical evidences nevertheless. It brought home one crucial point - if the recording source and/or the reproduction equipments has lower 'resolution' than 16bits, it is rather pointless to pursue 24bits digital recording.

Dynamic range in high fidelity systems

Reply #54 –
In practice, the situation is a little more complex than the basic theory suggests but even more weighted to the argument that for the consumer there is no point in higher than 16bit products.

The basic theory is that 16bit can encode a dynamic range of up to 96dB. This isn't the full story however! Virtually all modern digital recordings are made at 24bit (or 32bit float), the reason for this is to allow additional headroom during capture. Signal processing (mixing) is often carried out at 48bit, to avoid cumulative mathematical errors becoming audible. Once the mixing process is complete we usually have a master in 24bit format. During the mastering process the original 24bit digital audio files are reduced to 16bit for CD release. Part of this procedure is a process called Noise Shaped Dither. What this does is redistribute the 16bit quantisation noise into areas of the frequency spectrum where our hearing is least sensitive. Exactly where these frequency areas are and exactly how aggressive the redistribution of noise is, depends on the various different Noise Shaped Dither processors available and the settings employed. As a general rule, the noise is usually redistributed either above 12kHz and/or below 60Hz.

So what does all this mean? It means that by using Noise Shaped Dither the 96dB maximum dynamic range of a CD (16bit digital audio) can be increased perceptually to around 120dB. I say "perceptually" because the same amount of noise is still there, only now it's concentrated at the extremes of the human hearing spectrum where we are hundreds of times less sensitive to it. Indeed, using the most aggressive noise shaping dither, 16bit resolution can give a perceived dynamic range of up to 150dB in the most critical hearing band, the 1kHz-4kHz range. BTW, noise shaped dither is not a particularly new technology. It's been around well over a decade and you would be hard pressed to find a commercial CD released in the last 8 years which didn't employ it.

In conclusion, there are very good and justifiable reasons for using 24bit audio, but only for recording and mixing! For the consumer with a finished product, there are no advantages whatsoever, for 3 reasons: 1. No one in their right mind would master a recording using the full dynamic range available on CD. 2. If they tried, no one would have a system which could reproduce it and 3. If someone did use the full dynamic range and if a system could reproduce it, you would have to turn your amp up so high to hear the quiet parts that the loud parts would probably put you in a coma or even kill you!!

Some people have said, and indeed I myself have heard at times more dynamic range and higher quality sound on a 24bit DVD-A than on a CD version. This is due to differences in mastering and/or file handling, not to an intrinsic weakness in the 16bit format. I have to say, I am pretty annoyed at the marketing to consumers of 24bit audio and 24bit devices. The way digital audio works is fairly complex and most consumers simply aren't interested. Manufacturers of consumer equipment take advantage of this and suggest that 24bit audio has more data than than 16bit audio and is therefore more accurate and better. What they don't tell you is that 16bit is already beyond the ability of playback systems or indeed of the human ear! I remember one instance (about 4 years ago) where a website posted a 24bit and a 16bit version of the same section of music so consumers could download and hear the quality advantages of the 24bit version. It turns out that the 16bit version had been deliberately "doctored" to sound worse than the 24bit version. I doubt this was an isolated case!!

Cheers, G

Dynamic range in high fidelity systems

Reply #55 –
Quote from: Gregorio on
In practice, the situation is a little more complex than the basic theory suggests but even more weighted to the argument that for the consumer there is no point in higher than 16bit products.

What you say is correct (and has been pointed out many times before). I would only add that in most cases it is not even necessary to dither because the source material itself already has much less than 96 dB of dynamic range. First, it is difficult to capture sound with that much dynamic range (most don't even try). Second, these days lots of compression is added in the final stage, reducing dynamic range even further.

Dynamic range in high fidelity systems

Reply #56 –
Dithering anyway ensures to preserve the specific noise signature of the original recording at virtually no cost.

Dynamic range in high fidelity systems

Reply #57 –
Quote from: pdq on
What you say is correct (and has been pointed out many times before). I would only add that in most cases it is not even necessary to dither because the source material itself already has much less than 96 dB of dynamic range. First, it is difficult to capture sound with that much dynamic range (most don't even try). Second, these days lots of compression is added in the final stage, reducing dynamic range even further.


It is always necessary to dither when going from 24bit to 16bit! Dither effectively randomises the errors caused by the truncation process. This randomisation is heard in the reconstructed signal as noise, which is far preferable to the digital distortion caused by truncation without dither. This noise is what defines the dynamic range of digital audio bit resolutions and has been the accepted professional standard for bit reduction processes for more than 15 years. Noise shaping the dither is more recent but has still been around for more than a decade, as I said before.

If you're still not sure what the purpose of dither is and why it is always employed, please read this article by one of the most respected experts on digital audio:
http://www.cadenzarecording.com/Dither.html

Cheers, G

Dynamic range in high fidelity systems

Reply #59 –
Quote from: Gregorio on
It is always necessary to dither when going from 24bit to 16bit!
...
If you're still not sure what the purpose of dither is and why it is always employed, please read this article by one of the most respected experts on digital audio:
http://www.cadenzarecording.com/Dither.html


What PDQ meant is a sufficient noise level, that is already part of a signal, has the same effect as applying separate dither, in relation to the wanted signal. And that's true. The only practical difference between dithering and not dithering a signal, that is already subject to a lot of noise, is that the signal with dedicated dither applied will have a noise signature closer to the source's.

The article you have linked doesn't apply here, because, when noise free sine waves get truncated, far louder artifacts are created than when truncating a sine wave on top of a considerable noise floor (unshaped dither is basically not much else).

But as you I always dither, since I consider any noise, that is not quantization noise, part of the to be preserved signal.

Dynamic range in high fidelity systems

Reply #60 –
Quote from: googlebot on
I consider any noise, that is not quantization noise, part of the to be preserved signal.

Are you hoping to be able to decode an encrypted message or something?  Would that even work?  If the level of dither and quantization noise greater than the noise of the source, your original source noise would be lost (or at least part of it if you're using noise-shaped dither).  But who cares, it's just noise, right?

Dynamic range in high fidelity systems

Reply #61 –
Quote from: googlebot on
Dithering anyway ensures to preserve the specific noise signature of the original recording at virtually no cost.

True, but it does not increase the dynamic range, which was the subject under discussion.

Dynamic range in high fidelity systems

Reply #62 –
Quote from: greynol on
But who cares, it's just noise, right?


I tend not to agree. I have several favorite noises, others sound awful (does anyone like pink and white noise?). It was John Cage's intention in 1952 to bring exactly stuff like that to our attention, when he "composed" 4?33?. A comparison of a vintage recording to a 90's version to perfect digital silence is clearly ABXable (TOS 8 ) and that* is part of the point he was trying to make.

Anyway, wether one find's it odd to compare recorded silence or not, it should be left to the listener to decide and not the engineer, who just needs to check a box to preserve as much as possible.

* Besides the sound signature of the room, where it was performed, its contained objects, and persons.

Dynamic range in high fidelity systems

Reply #63 –
Quote from: greynol on
If the last 8 bits of the 24-bit source are nothing but noise then dithering is not necessary.  A perfect example would be the digitization of vinyl.


Quoted from the article I posted a link to above: "The bottom line is that WHENEVER a signal goes from a higher resolution to a lower resolution it is necessary to dither in order to avoid the artifacts provided by truncation that have been shown above. This means that whenever signals go from 48 bit resolution for processing to 24 bit resolution, or 24 bit resolution for mixing to 16 bit, or even analog (infinite resolution) to 24 bit during A/D conversion dithering needs to happen."

Look at figure 10 of Nika's article and notice the quantisation distortion peaks from truncation, some of which are 18dB higher in the truncated 16bit file than in the dithered version (and that is with a noisy original recording). Dither on the other hand is a specific TDPF (or a coloured relative when noise shaped), which statistically randomises all quantisation errors, unlike noise captured during recording which may correlate with the program material and cause harmonic distortions.

PDQ - "True, but it does not increase the dynamic range, which was the subject under discussion."

Yes it does, read the article and go down to figure 15!!

googlebot - "But as you I always dither, since I consider any noise, that is not quantization noise, part of the to be preserved signal."

Exactly. In Audio Post for example, what to most people is just noise is in fact called Room Tone and a great deal of time and effort is often put into recording, mixing and reproducing Room Tones. No one would be happy if these room tones were obliterated with quantisation distortion! Maybe in some genres of music (electronica for example) all noise is unwanted but for classical music in particular this is certainly not the case.

I'm not the dither police, I'm just stating that professional practice (for good reason) is to dither, if someone wants to truncate without dither that's their lookout.


Cheers, G

Dynamic range in high fidelity systems

Reply #64 –
Quote from: Gregorio on
PDQ - "True, but it does not increase the dynamic range, which was the subject under discussion."

Yes it does, read the article and go down to figure 15!!

Figure 15 appears to show the benefit of applying dither to a sine wave, which we all know increases the dynamic range. Where is the plot that shows that adding dither to a noisy signal increases the dynamic range over truncation?

Dynamic range in high fidelity systems

Reply #65 –
Quote from: Gregorio on
Quoted from the article I posted a link to above: "The bottom line is that WHENEVER a signal goes from a higher resolution to a lower resolution it is necessary to dither in order to avoid the artifacts provided by truncation that have been shown above. This means that whenever signals go from 48 bit resolution for processing to 24 bit resolution, or 24 bit resolution for mixing to 16 bit, or even analog (infinite resolution) to 24 bit during A/D conversion dithering needs to happen."

So despite all that I've read, there is no such thing as a self-dithered signal?

Dynamic range in high fidelity systems

Reply #66 –
Dither

Paul

     
"Reality is merely an illusion, albeit a very persistent one." Albert Einstein

Dynamic range in high fidelity systems

Reply #67 –
Quote from: greynol on
Quote from: Gregorio on
Quoted from the article I posted a link to above: "The bottom line is that WHENEVER a signal goes from a higher resolution to a lower resolution it is necessary to dither in order to avoid the artifacts provided by truncation that have been shown above. This means that whenever signals go from 48 bit resolution for processing to 24 bit resolution, or 24 bit resolution for mixing to 16 bit, or even analog (infinite resolution) to 24 bit during A/D conversion dithering needs to happen."

So despite all that I've read, there is no such thing as a self-dithered signal?


As common as self-dithered signals are,  it is unwise to stake your algorithm on always having one.

Dynamic range in high fidelity systems

Reply #68 –
Quote from: pdq on
Figure 15 appears to show the benefit of applying dither to a sine wave, which we all know increases the dynamic range. Where is the plot that shows that adding dither to a noisy signal increases the dynamic range over truncation?


Anything recorded in digital audio is a sine wave or combination of sine waves. The Nyquist-Shannon theory only applies to sine waves. So even noise on the recording is just a combination of sine waves. Also, all digital recordings have noise, particularly when recording in 24bit, which has a theoretical noise floor way beyond the practical dynamic range of any recording chain. Even in a world class studio, it is extremely unlikely that a track can be recorded in which any of the 8 LSBs (in a 24bit recording) contain anything other than noise.

Figure 15 doesn't show that dither increases dynamic range it shows that noise-shaped dither increases dynamic range over standard TDPF dither. In fact, dither does not, as you state, increase dynamic range, it reduces dynamic range! This reduction of dynamic range is a trade-off. It's better to have a reduced dynamic range than to have higher peaks of enharmonic quantisation distortion (as clearly demonstrated in Figure 10 of the article). The increase in dynamic range comes from noise-shaping the dither. It's pretty difficult to find any detailed graphs comparing dynamic range truncation to noise-shaped dither bit reduction. Although there are plenty comparing TDPF dither with noise-shaped dither. This is because there is simply no point, applying dither to a bit reduction procedure rather than just truncating is now so widely accepted professionally as to be considered an axiom of professional digital audio processing. However, have a look at this document:
http://www.meridian-audio.com/ara/coding2.pdf
In particular "IN-BAND NOISE SHAPING AND PRE-EMPHASIS" page 7 and the graph "Figure 19" on page 31.

greynol: "So despite all that I've read, there is no such thing as a self-dithered signal?"

Yes there is but it is not applicable to the reduction of a 24bit recording to a 16bit recording. During the initial ADC process, dither has to be applied to a 24bit result. The theoretical noise floor of 24bit is -144dB. The noise caused by electrons colliding inside a single resistor is higher than that, which of course results in the fact that the practical noise floor of a 24bit converter is nowhere near the theoretical noise floor of 24bit digital audio. In practice therefore, there is no such thing as a 24bit converter, sure a 24bit converter can output or input a 24bit format file but no converter can actually resolve more than about 20 bits of resolution. For this reason, many 24bit converters handled dither purely by utilising the noise within it's own electronic circuitry, hence self-dither. However, as with the noise inherent in a recording, this noise is not necessarily decorrelated from the actual signal and therefore enharmonic quantisation distortion peaks can still occur. For this reason, most modern higher quality 24bit ADCs now apply some form of TDPF dither rather than rely on self-dithering. However, these points are not particularly relevant to our discussion as dithering (or self-dithering) a 24bit signal to avoid enharmonic quantisation distortion peaks at say the -125dB level is insignificant compared to dithering a 16bit signal where otherwise quantisation distortion could peak in the -70dB to -80dB region (and therefore be audible!).

Cheers, G

Dynamic range in high fidelity systems

Reply #69 –
Quote from: Gregorio on
greynol: "So despite all that I've read, there is no such thing as a self-dithered signal?"

Yes there is but it is not applicable to the reduction of a 24bit recording to a 16bit recording.


It is if the original noise is greater than 16-bit dither.


I've yet to see a rigorous mathematical proof of "self dithering" (though there may be one somewhere) - people just tend to look at the residual, and try to spot anything that's not white noise.

Then there's lossyWAV

Cheers,
David.

Dynamic range in high fidelity systems

Reply #70 –
Quote from: Gregorio on
During the initial ADC process, dither has to be applied to a 24bit result.

You are probably thinking of the old-fashioned sample-and-hold type ADC, which did indeed benefit from added dither.

As far as I know, modern-day sigma delta ADCs don't need or benefit from dither.

Besides, as you say, the noise present in any real audio source is far greater than -144 dB, so adding -144 dB of dither would make a negligible difference.

Dynamic range in high fidelity systems

Reply #71 –
This touches the edge of cryptography. The benefit of dither is that you have full knowledge about its pseudo-random properties. You don't have that about self-dither. Self-dither may even contain more true random-ness than a pseudo-random sequence, but still its actual sequence may have a much higher regularity. Imagine a case where a sequence is mostly eight 1s followed by two zeros, sometimes (seeded from a truly random event as electrons colliding) it is 8 ones followed by three zeroes. The latter can be made a even "better" random sequence than the pseudo-random variant by normalization, but as it is it doesn't have to be sufficient just because it is seeded by true randomness and uncorrelated to a signal.

Dynamic range in high fidelity systems

Reply #72 –
Quote from: 2Bdecided on
Quote from: Gregorio on
greynol: "So despite all that I've read, there is no such thing as a self-dithered signal?"

Yes there is but it is not applicable to the reduction of a 24bit recording to a 16bit recording.


It is if the original noise is greater than 16-bit dither.


My position is still "no it's not". First of all, what is "noise"? Is it just electronic interference from the recording chain/mixing process or is it the ambience and noise floor of the recording environment? Probably it's some mixture of both. So this raises serious issues.

Firstly, this noise is unlikely to be truly random, it might sound like white noise but is it really? Look again at Figures 3 and 4 in the article I posted a link to (http://www.cadenzarecording.com/Dither.html). The greatest quantisation harmonic distortion due to truncation is happening in the frequency range which is filled only with electronic interference not at the sine wave frequency of 100Hz. If it's not truly white noise, a TDPF or a statistical relative of TDPF then it's likely that correlation with the program material will occur, potentially resulting in audible harmonic distortions as evidenced in Figure 4 of the article. Under these conditions the noise is not acting as a dithering agent but purely as program material (wanted or unwanted). Secondly, applying a noise shaped dither will retain the harmonic integrity of the original signal, including the noise (both wanted and unwanted). Therefore whatever constitutes the noise, including the ambience and noise floor of the original recording will also be preserved.

In short, there is no way to know if the noise on a recording will cause "self-dithering" or will contribute to harmonic distortion, so to be safe, it's always better to apply proper dither when performing bit reduction from 24bit to 16bit. Seems like I'm back to where I started! 

pdq - "You are probably thinking of the old-fashioned sample-and-hold type ADC, which did indeed benefit from added dither. As far as I know, modern-day sigma delta ADCs don't need or benefit from dither. Besides, as you say, the noise present in any real audio source is far greater than -144 dB, so adding -144 dB of dither would make a negligible difference."

AFAIK, modern 24bit converters oversample, sometimes in excess of 11mFs/s, using several bits per sample. This primary quantisation process in then decimated down to the sample frequency/bit depth set for output. I'm reasonably certain that TDPF dither is applied (by the better manufacturers) at least during the decimation process if not during the original sampling. Bare in mind that dither wouldn't be added at -144dB but at about -136dB. The amount of dither added is usually equivalent to a little more than the value of the LSB (6dB). However, without dither, quantisation distortion could peak as high as the -120dB range, still pretty insignificant I grant you but potentially enough of a problem that higher quality ADC makers use TDPF dither. Incidentally, in the professional audio world, dither is sometimes explicitly employed when going from a 48bit signal to a 24bit signal, this is to avoid potentially audible artefacts during the mixing/summing process. My point here is that if dither is sometimes expressly required at 24bit it is definitely required at 16bit, where everything related to the noise floor and quantisation errors is potentially 48dB louder!

Cheers, G

Dynamic range in high fidelity systems

Reply #73 –
Quote from: pdq on
Quote from: Gregorio on
During the initial ADC process, dither has to be applied to a 24bit result.

You are probably thinking of the old-fashioned sample-and-hold type ADC, which did indeed benefit from added dither.

As far as I know, modern-day sigma delta ADCs don't need or benefit from dither.


As I understand it, they randomize quantication error as an inherent part of their operation.

Quote
Besides, as you say, the noise present in any real audio source is far greater than -144 dB, so adding -144 dB of dither would make a negligible difference.


Since virtually every good audio ADC is sigma-delta, they all benefit from the inherent randomization of quantization error and thus as you said, need no additional dither.

Dynamic range in high fidelity systems

Reply #74 –
Quote from: Gregorio on
it's always better to apply proper dither when performing bit reduction from 24bit to 16bit. Seems like I'm back to where I started!

Realizing I sound like a pedant, you originally said "necessary" not "better".  As far as this forum is concerned, whether it is better (let alone necessary!) in any given instance must be determined by a double blind test and not by graphs.

Feel free to give us two samples of digitized vinyl that has been recorded at 24 bits, de-clicked, then normalized and then converted to 16 bits, one through truncation and one through the dithering of your choice as well as the ABX results showing you can hear the difference.