Dithering question

[mshilarious]

OK, let's just try this:

Truncated Organ

These are three organ pipes, of increasing pitch and thus decreasing quantity of overtones (due to bandwidth limitation at A/D). Recorded at 24 bit with a KSM44 at 12 inches, not exactly a natural way to record pipe organ--these are mouthblown. There is no way an installed pipe in a pipe organ could produce as noise-free a recording (blower noise, room noise)--and this recording isn't a pure series of sine waves, because moving air produces noise. Then I applied a fade to each pipe, then truncated to 16 bit.

Anyway. Please be aware that especially the last pipe will be very, very painful to listen to on headphones at a calibrated monitoring level. You'll basically be feeding an 85dBSPL 3kHz sine wave directly into your brain.

Now, can I measure QD on any of these samples? Yes, I can, the last second of the fade of the last pipe shows some QD at -140dBFS. You don''t need young ears to hear it, the QD is not at a high pitch. But I did need to increase gain +14dB above my reference point to hear it, that would make the beginning of that pipe nearly 100dBSPL, which hurts real bad. On monitors, I believe it would be impossible to hear without a stupid amount of gain above reference level, because of the masking effects of room noise--you would probably need to set your monitors to 120dBSPL.

My conclusion--if you are recording something like a solo piccolo-ist who isn't breathy (or the aforementioned pure synth stuff), consider dither. Otherwise, get back to selecting that bus compressor . . .

 

[noisewreck]

The "no_dithered file" is truncated to 7 bit and the "dithered" is, of course, dithered then truncated to 7 bit...[/B]

I haven't listened to the examples and I am not going to listen to...

You lost me at 7 bits.

I can start hearing quantization noise at 12 bits on most music. But at 16 bits? Especially since most modern music is squashed into a 3db of dynamic range and pressed right against 0dBFS? Can you really hear the last 8 bits in action? Or any noise for that matter?

Blah.

Meh... whatever. I don't really care. I internally process at 32 bits and then dither to 16. Just so that I don't have to worry about all this nonsense.

 

[masteringhouse]

But Ethan's point is that dither is unnecessary to prevent quantization distortion on any real-world acoustic signal, and he's correct.

Actually dither is there for the very reason of helping to prevent the audbile effects (or masking) quantization distortion. It also adds up given a greater number of tracks.

See the following for examples:http://www.masteringhouse.com/dither/dither.html

I can't say whether a given person can hear it or not, this is something that you need to determine for yourself given a good monitoring system.

Some people think Adam Sandler is funny too.

 

[mshilarious]

Actually dither is there for the very reason of helping to prevent the audbile effects (or masking) quantization distortion. It also adds up given a greater number of tracks.

See the following for examples:http://www.masteringhouse.com/dither/dither.html

I can't say whether a given person can hear it or not, this is something that you need to determine for yourself given a good monitoring system.

Some people think Adam Sandler is funny too.

Tom, did you actually read anything I wrote, or is that a stock response? Your link is another test tone, not a real-world acoustic recording.

I'll repeat: any real-world acoustic source is sufficiently complex and/or noisy as to effectively self-dither. YOU DON'T NEED TO ADD NOISE, there is already enough!

If you can prove otherwise, please do. No 8-bit truncations please, unless your mastering house specializes in producing audio for 1980s-vintage video games . . .

Let's dispense with boasting about how great your monitoring chain is. When we are talking about audibility of low-level signals, headphones win every time. Every time! Do audiologists come to your mastering house to conduct hearing tests? I'm guessing not. I don't care how quiet your studio is, headphones are quieter. You don't need a super-accurate monitoring chain to hear a distortion product that lives at 1kHz, do you? I should hope not!

Anyway, this is all irrelevant, because in my organ sample, I can measure a distortion product . . . at -140dBFS. I can't hear it at standard calibrating listening levels on monitors (which I built myself, thank you very much), nor on headphones. It's not until I turned the headphones up an extra 14dB that it even meets the barest criteria for audibility.

But I can measure it. Easily. So if I can measure it, then I don't need to hear it, do I? I know that my ability to measure is at least 14dB better than my ability to hear (actually, it's closer to 40dB).

So let's have your review of my listening sample, or your own experiment with a real source.

 

[SouthSIDE Glen]

one must accept that the added noise of dither (or truncation) is the selection between different effects.

That is exactly what it is.

Quantization error is often referred to as being "introduced" by truncation and being "neutralized" by dither. These descriptions can be mislead one to the wrong picture.

In fact, both truncation and dither ARE quantization error themselves. They are just different flavors and styles of error.

It can be mathematically and experimentally demonstrated that dither (in one form or another) can indeed present a more seemingly "natural" image or phychoacoustically translatable image to the brain than truncation alone, and, you are right, Jon, that the simpler, more "predictable" the original waveform, the more likely that dither will "help".

This is all fine in the laboratory. The problem arises when you start introducing other real-world variables into the "experiment" such as signal the bit depth of the dithering, signal dynamic range, signal complexity, signal-to-noise ratio, signal playback quality, signal playback volume, hearing quality of the listener, and personal tastes of the listener, the effects of 16-bit dithering tend to get lost in the fog of reality.

Can some folks hear some things in some recordings that are perhaps not easily measurable by instrument because they are more of a cumulative perception thing than a sampled measurement thing, and do some people have a better ear than others for such things? I believe so. This is the 0.1% (or so, that's not meant as hard data ) of instances that keeps me from saying that 16-bit dither is completely inaudible.

But there's a difference between audible and meaningful. Take enough "golden ears" and ask them to listen to a set of real-world recordings that are not A/B comparisons of dither, and ask them the question, "Does this recording have dither in it?", and I guarantee you that they will average a random distribution of 50% correct answers. There may be a small hand full of auditory savants out there (maybe one in 500 million) that have the seemingly supernatural ability to hear 16-bit dither as it's own sound regardless of the source material and without requiring a non-dither comparison, but mixing for them is about as useful as composing and arranging for those who can hear (not feel, hear) 22kHz dog whistles.

G.

 

[mshilarious]

That is exactly what it is.

Quantization error is often referred to as being "introduced" by truncation and being "neutralized" by dither. These descriptions can be mislead one to the wrong picture.

In fact, both truncation and dither ARE quantization error themselves. They are just different flavors and styles of error.

100% correct! The goal of dither is to de-correlate the error such that it is distributed across the spectrum instead of existing as very correlated distortion.

But if the signal in question is sufficient to de-correlate the error all by itself, then all dither does is change the flavor of the de-correlated error.

 

[masteringhouse]

Tom, did you actually read anything I wrote, or is that a stock response? Your link is another test tone, not a real-world acoustic recording.

I'll repeat: any real-world acoustic source is sufficiently complex and/or noisy as to effectively self-dither. YOU DON'T NEED TO ADD NOISE, there is already enough!

If you can prove otherwise, please do. No 8-bit truncations please, unless your mastering house specializes in producing audio for 1980s-vintage video games . . .

Let's dispense with boasting about how great your monitoring chain is. When we are talking about audibility of low-level signals, headphones win every time. Every time! Do audiologists come to your mastering house to conduct hearing tests? I'm guessing not. I don't care how quiet your studio is, headphones are quieter. You don't need a super-accurate monitoring chain to hear a distortion product that lives at 1kHz, do you? I should hope not!

Anyway, this is all irrelevant, because in my organ sample, I can measure a distortion product . . . at -140dBFS. I can't hear it at standard calibrating listening levels on monitors (which I built myself, thank you very much), nor on headphones. It's not until I turned the headphones up an extra 14dB that it even meets the barest criteria for audibility.

But I can measure it. Easily. So if I can measure it, then I don't need to hear it, do I? I know that my ability to measure is at least 14dB better than my ability to hear (actually, it's closer to 40dB).

So let's have your review of my listening sample, or your own experiment with a real source.

I don't recall ever boasting about my monitoring chain, have you read my response? I was simply pointing out that saying "dither is unnecessary to prevent quantization distortion on any real-world acoustic signal" is fundamentally flawed. You can argue this all you want and you will still be wrong. Have you seen the graphs and heard my examples? Granted these are overblown to a point, but it's to make them audible and obvious.

The issue everyone brings up is can you hear it? This is subjective and you can argue that point (and we have on this forum) until you're fingers bleed. To say that something exists or doesn't exist because any one group of people can't hear it approaches religion, not science or fact.

General noise is not distributed in the same way as dither, nor does it provide the same function let alone comparing it to how noise shaped dithers work. Noise generally masks (or sweeps the issue under the rug), while dither attempts to address the problem directly. Dither helps distrubute quantization distortion (see my graphs) so that it is less audible, but adds another type of noise, just hopefully not a cold and grainy sounding. Noise just covers it up like any other sound, like throwing mud over a rust spot on your car to hide the blemishes.

Anyway I'm not going to get into this again, been through that.

 

[mshilarious]

I don't recall ever boasting about my monitoring chain, have you read my response?

Yes, I read your entire response and your web page. You are using a 500Hz sine wave. That is not a real-world acoustic source, as I stipulated.

As for boasting:

I can't say whether a given person can hear it or not, this is something that you need to determine for yourself given a good monitoring system.

Some people think Adam Sandler is funny too.

In other words, Adam Sandler sucks and so does my monitoring chain. Let's not be coy, shall we?

Anyway, my FFT works fine, as does yours. FFT my organ sample, please, compare the distortion with the fundamental signal peak at -20dBFS and -80dBFS, as in your sine wave sample. Is it comparable to your 500Hz sine wave example? Why or why not?

I was simply pointing out that saying "dither is unnecessary to prevent quantization distortion on any real-world acoustic signal" is fundamentally flawed. You can argue this all you want and you will still be wrong. Have you seen the graphs and heard my examples? Granted these are overblown to a point, but it's to make them audible and obvious.

Yes, but THOSE AREN'T REAL WORLD ACOUSTIC SOURCES. Do you master a lot of single sine waves? I mean single sine wave noise-free sources, not sine waves in the sense of Fourier's theorem.

The issue everyone brings up is can you hear it? This is subjective and you can argue that point (and we have on this forum) until you're fingers bleed. To say that something exists or doesn't exist because any one group of people can't hear it approaches religion, not science or fact.

That's why I didn't just listen, I measured. As I stated, my ability to measure exceeds my ability to hear by at least 40dB. That is science, man

General noise is not distributed in the same way as dither, nor does it provide the same function let alone comparing it to how noise shaped dithers work. Noise generally masks (or sweeps the issue under the rug), while dither attempts to address the problem directly. Dither helps distrubute quantization distortion (see my graphs) so that it is less audible, but adds another type of noise, just hopefully not a cold and grainy sounding. Noise just covers it up like any other sound, like throwing mud over a rust spot on your car to hide the blemishes.

Not exactly. Your basic ol' white noise works just fine as dither, so long as it's applied before truncation! I can demonstrate that ad nauseum on sine waves if you like. The problem is white noise is more audible than necessary, so we like to use noise-shaped dither instead, that increases dynamic range across the more audible parts of the spectrum vs. the white noise.

That's a great theory, but when your signal only has a dynamic range of 80dB, it doesn't matter what you do to the LSB, you can't increase dynamic range one bit (sorry, bad pun). You get the noise you start with--but you also get no QD, because that noise that already lives in the last 3 bits is doing the same job the dither would.

Anyway I'm not going to get into this again, been through that.

I don't think you have. Try my experiment for yourself. Show me an FFT of reasonably audible QD on a real-world acoustic source. Believe me, I wanted to do that very thing, I tried pretty hard . . . and I failed.

 

[masteringhouse]

Not exactly. Your basic ol' white noise works just fine as dither, so long as it's applied before truncation! I can demonstrate that ad nauseum on sine waves if you like. The problem is white noise is more audible than necessary, so we like to use noise-shaped dither instead, that increases dynamic range across the more audible parts of the spectrum vs. the white noise.

That's a great theory, but when your signal only has a dynamic range of 80dB, it doesn't matter what you do to the LSB, you can't increase dynamic range one bit (sorry, bad pun). You get the noise you start with--but you also get no QD, because that noise that already lives in the last 3 bits is doing the same job the dither would.

You increase dynamic range whenever you go to a longer wordlength. Many plug-ins do this under the covers, in addition they are adding more quantization distortion which accumulates to a level proportional to the number of times you process. Even something as simple as a volume adjustment adds a degree of quantization distortion. Given enough of these operations I believe it can be audible to most, or at the very least your are increasing your odds of this.

Anyway don't believe me I really don't care. This article explains all of this well (including the benefits of proper dither over simple white noise).

http://www.users.qwest.net/~volt42/cadenzarecording/DitherExplained.pdf

 

[BushmasterM4]

Ok. When rendering your 24 bit tracks or track down to 16bit, should you use 16bit or 24bit dithering ?

 

[SouthSIDE Glen]

The issue everyone brings up is can you hear it? This is subjective and you can argue that point (and we have on this forum) until you're fingers bleed. To say that something exists or doesn't exist because any one group of people can't hear it approaches religion, not science or fact.

(BTW, Good to see you again Tom; you've been avoiding us little people lately )

I see it as just the opposite. To say that something that has little if any measurable or humanly detectable real-world effect is something that is important to consider sounds much more like faith than real-world science to me. It's not like the butterfly effect applies here.

People can argue theory all they want and provide all the laboratory experiments they want to prove a proprietary point), but for me, here's the beginning and the end of the story:

When it comes to mastering, if I am giving my work to a consummate pro such as yourself, I expect my ME to *try* dithering, if possible choosing an algorithm or two based upon what they hear in the program material and their historical knowledge of what they thought had worked best in previous similar situations.

When they try it, they need to ask themselves two questions:

1) Is the result of dithering even audible to me on this project?

2) Does any audible difference caused by the dithering positively affect the material in a way as desired or called for by the producer/client first and the material second?

What I expect from my ME at that point is honesty; i.e. that if the answer to either of those questions is in the negative, that they not only remove the dithering, but admit to the client and themselves that the dither was not part of the final product because it was not a proper change under the circumstances; and that if they did dither, that there was a legitimate reason for it as stipulated above, and not just because that's what other MEs do.

If one is a home wrecker doing their own mastering, actually, I would advise nothing different. Blind faith in *any* effect or process without judging every situation by it's merits is not engineering, it's religion.

If dithering makes it sound better, fine. Use it. Knock yourself out. I'm happy, you're happy.

It it sounds worse to you, don't use it. Period. Trust your ears, not a blind faith.

If you can't hear the difference, don't do it. Not only does this follow a general rule of thumb of keep your processing as short as necessary, don't flip any bits that don't need to be flipped. But if there is a difference and you just can't hear it yourself, there's no guarantee whether that difference is actually a good one or not. You HAVE TO hear it to know.

Screw the theory, screw the faith, screw the pride, use your ears and be honest about it, and it can be amazing how the debate will disappear and the answer will reveal itself.

G.

 

[NYMorningstar]

(NUTSHELL)I'll repeat: any real-world acoustic source is sufficiently complex and/or noisy as to effectively self-dither. YOU DON'T NEED TO ADD NOISE, there is already enough!(/NUTSHELL)

Thank you!

 

[Ethan Winer]

Based on what everyone is writing, I’m leaning towards not doing it.

I don't bother with dither myself in SONAR, but I don't tell people not to use dither because it's free and takes no longer to render a mix. My only interest is in explaining the facts and science of audio. The fact is, dither is not audible unless the music gets down to like -40 dB, and the volume is cranked enough to make that -40 dB signal reasonably loud.

If dither really was audible on typical pop music, I'd have received half a dozen correct responses to my Dither article linked above. In all the time since that article has been online, I have never received even one response. And not because nobody has read the article. I just checked, and it's gotten 1,475 hits this year so far.

--Ethan

 

[masteringhouse]

Hey G.

I'm not really in total disagreement with all that you said except that one has to ask is there something there that I may not be hearing today for a variety of reasons? It may be that one doesn't have the experience (think about how long it takes for many newcomers to hear the overuse of compression), maybe it's a fault with a monitoring system, room, or maybe one is just having a bad day. Perception is not the same for everyone obviously so you have to go beyond your own experience if you plan to mass produce something. At times, FFTs or other tools can help in this area, though I wouldn't want to use one for making a decision regarding how I want to shape an EQ.

I'm not trying to be pretentious when I say that I hear a difference not only between truncation and dithered material, but also between different types of dithering. Yeah it's subtle, but subtlety adds up to give us that last 20%. Try different noise shaping algorithms in your favorite limiter and see if you can't hear a difference. If it allows it, also try turning dithering on and off and listen carefully. I would be curious to see responses from people can hear absolutely no difference doing this and those that can.

I don't make my money selling dithering software andhave no hidden agenda regarding dithering. If I really didn't feel that it mattered I wouldn't even bother with these posts. I just feel that it's important for people to know the facts and they can perceive whatever they like from them.

BTW one last comment. Dithering is more than adding noise, it helps to remove harmonic distortion caused by quantization. Granted on some instruments rich with harmonics it may be more difficult to hear, you have to listen to the change in timbre of an instrument. That's one reason why people use sine waves to test this. It's a pure tone and much easier to hear this effect.

 

[Ethan Winer]

the audbile effects (or masking) quantization distortion ... adds up given a greater number of tracks.

DAW software dithers the entire mix when it's exported. So even if dither were audible, I can't see what the number of tracks has to do with anything. A complex mix could come from 35 separate tracks, or it could be from 35 microphones in a live-to-two-track session.

It's also worth repeating that most mixes have a noise floor much higher than dither which is down around -90 dB. So whether the worry is truncation distortion or the noise added by dither, both go out the window when the residual noise is 10 or even 20 dB louder.

Added after you seeing your last post above: Any distortion added by not using dither is also -90 dB, so it too is inaudible on "typical pop music" that's mixed close to 0 dB.

--Ethan

 

[masteringhouse]

DAW software dithers the entire mix when it's exported. So even if dither were audible, I can't see what the number of tracks has to do with anything. A complex mix could come from 35 separate tracks, or it could be from 35 microphones in a live-to-two-track session.

These are two different things though aren't they. Lets forget about dither and add 2 tracks of noise versus 35 tracks of noise. When summed together which has the higher noise floor?

It's also worth repeating that most mixes have a noise floor much higher than dither which is down around -90 dB. So whether the worry is truncation distortion or the noise added by dither, both go out the window when the residual noise is 10 or even 20 dB louder.

Most of the subtlety of music goes out the window in a noisy environment. Take for example listening in a car. Does that mean that we shouldn't be concerned about what's "down there"? And more than what's down at the lower noise floor it's my belief that quantization distortion can be heard even above -90dBFS.

Added after you seeing your last post above: Any distortion added by not using dither is also -90 dB, so it too is inaudible on "typical pop music" that's mixed close to 0 dB.--Ethan

Is it? If so why do so many other people hear the differences in noise shaped dithering algorithms that difference is also below -90. And I'm not talking about "golden ears"' but about about clients that come here to the studio.

Also see:
http://www.24-96.net/dither/

 

[mshilarious]

I'm not trying to be pretentious when I say that I hear a difference not only between truncation and dithered material, but also between different types of dithering. Yeah it's subtle, but subtlety adds up to give us that last 20%. Try different noise shaping algorithms in your favorite limiter and see if you can't hear a difference. If it allows it, also try turning dithering on and off and listen carefully. I would be curious to see responses from people can hear absolutely no difference doing this and those that can.

I have never argued that truncation vs. various types of dither shouldn't sound different, and they should sound different from the original 24 bit wav. I'm not arguing that!

What I am arguing is that it's not because of QD, it's because the flavor of noise is different.

Granted on some instruments rich with harmonics it may be more difficult to hear, you have to listen to the change in timbre of an instrument. That's one reason why people use sine waves to test this. It's a pure tone and much easier to hear this effect.

It's not that it's much easier to hear because the source is simple, it's that the distortion is much worse for the sole reason that the source is simple. QD is harder to hear on a complex source because there is much less of it!


Let's try some different analyses--here are a series of FFTs on various sources. One technique I have used for some of these is a null test of a truncated file against the 16 bit original. The remainder is the quantization error, so we can investigate the character of that error. You can see in most cases, it is a greatly attenuated version of the original. Some graphs show an increase in harmonic distortion, note though that is indeed harmonic distortion and not the sort of inharmonic stuff you get with QD on noise-free sine waves.

Tests:

- a null test on one of the organ notes--only an attenuation of the original signal remains, along with broadband noise.

- a null test on the organ note, dithered--unsurprisingly, all that remains is the dither (UV22HR in this case).

- a truncated bass note. I played a single note (Bb) on a bass, with the tone control full off, gently using my thumb, and letting it naturally fade out. The bass was DI'ed into a quiet preamp, so there is no acoustic noise here, and relatively little electrical noise. The goal is to generate a minimum of natural harmonics and keep all noise sources as low as possible, so we can try to replicate the hideous QD you will see on a noise-free sine wave source.

This is different from the acoustic organ pipe, so we can see QD peaks a bit higher--still all harmonic distortion though, and with the natural second harmonic down to -50dBFS (this is near the end of the fadeout, worst case scenario), the added distortion lives at -120dBFS--0.0001% THD compared with 0dBFS, 0.03% THD compared with the -50dBFS harmonic. That latter figure sounds bad, but it is only relevant if you add an egregious amount of gain after truncation, and why would you do that? That's just a bad engineering practice.

This is pretty much worst-case scenario for QD on a non-synthetic source, I would have to think. I'll upload the .wavs if anybody wants to try and hear the difference.

- Same test, but on a fadeout of a low-frequency acoustic source--a didgeridoo. Note the fundamental level: -90dBFS. Try that with a noise-free sine wave, the result is horrible. But here, no apparent QD at all, just the same noise at either 16 or 24 bit (they are both there, try to see the difference!) Conclusion: more complicated sources suffer from less QD as a function of their complexity.

- Last test: a null of a drum overhead 16 bit trunc vs. 24 bit. This is the worst snapshot in the whole clip, much of it is flat white noise. The little peaks you see sound like a really quiet cymbal crash bleeding through the noise, and in fact that's what it is. It's very hard to analyze a cymbal because it is not a harmonic event. But I don't hear any hideous distortion in the nulled file, just a wash so quiet you'd have to use headphones or really crank your mains to hear it.

My conclusion is that QE on a real-world source is expressed almost entirely as noise, very secondarily harmonic distortion, and probably not at all as the inharmonic distortion you see on noise-free sine waves. The harmonic distortion doesn't seem exist on real sources much above -120dBFS, and likely much lower on a typical track.

 

[mshilarious]

And the drum picture:

 

[masteringhouse]

I have never argued that truncation vs. various types of dither shouldn't sound different, and they should sound different from the original 24 bit wav. I'm not arguing that!

What I am arguing is that it's not because of QD, it's because the flavor of noise is different.

I won't disagree, that was not the point I was trying to make. The fact that you can hear differences in types of noise shaping below -90 dBFS even with audio far above that level shows that artifacts from truncation and dithering are in fact audible.

Nice graphs, at some point when I have time I will have to try something similar here.

Best,
Tom

 

[mshilarious]

I won't disagree, that was not the point I was trying to make. The fact that you can hear differences in types of noise shaping below -90 dBFS even with audio far above that level shows that artifacts from truncation and dithering are in fact audible.

Yes, I agree. And I agree with Glen in that it's incumbent upon the ME to try different types of dither and yes, sometimes truncation, and make a qualitative analysis.

What's happened to date is a knee-jerk response based upon a mathematical law which has not been fully described in such analyses, and that's not always the appropriate engineering decision.

It may seem semantic, but I see it as part of a larger debate that too much of the pro and especially semi-pro market doesn't want to have. We know that noise and distortion and the character thereof have a strong impression on our sensory experience of sound, but too many people out there don't want to have the conservation on that basis. Science can tell us what is technically better, or simply what is different, but subjective analysis is required to determine what sounds best to a majority or even an identifiable subset of listeners.

The trouble is because one can't scientifically determine what sounds best, it is thus concluded that science cannot aid us in understanding what qualitative selections we have made, and how to facilitate that process, and that's really false.