http://dspace.cc.tut.fi/dpub/bitstream/handle/123456789/127/aumala.pdf?sequence=1
shows systematic errors of various dithers as a function of dither amplitude vs lsb size
the Gaussian plot is fairly convincingly near 0 for stdev > ~ 0.6 lsb
I looked up some of these refs while being forced to use the 12 bit integrated ADC on a uC for a product that really wanted ~14 bit resolution/noise, it was fast enough for 8x or more oversampling but just not linear enough for better than 13 bits even with dither/averaging
since product volume wasn't likely to reach 1K units/yr we could have saved enough in development, coding, testing time for a decade's production "extra" part cost by just using an external 16 ADC chip but the boss was convinced I was “Gold Plating” the product
I considered subtractive dither added to the ADC input per the above paper, did have to use a complicated sample "scrambling" multiplexing scheme when it turned out the internal ADC/mux had additional mux "pattern" noise that was invisible at 2^n interleave schemes but bit us badly with our 6 channels
shows systematic errors of various dithers as a function of dither amplitude vs lsb size
the Gaussian plot is fairly convincingly near 0 for stdev > ~ 0.6 lsb
I looked up some of these refs while being forced to use the 12 bit integrated ADC on a uC for a product that really wanted ~14 bit resolution/noise, it was fast enough for 8x or more oversampling but just not linear enough for better than 13 bits even with dither/averaging
since product volume wasn't likely to reach 1K units/yr we could have saved enough in development, coding, testing time for a decade's production "extra" part cost by just using an external 16 ADC chip but the boss was convinced I was “Gold Plating” the product
I considered subtractive dither added to the ADC input per the above paper, did have to use a complicated sample "scrambling" multiplexing scheme when it turned out the internal ADC/mux had additional mux "pattern" noise that was invisible at 2^n interleave schemes but bit us badly with our 6 channels
Last edited:
Sy,
I could not locate the original experiment I was referring to, but came across some like experiments in Blauert on p. 153, table 2.3 "Survey of measurements of lateralization blur in the median plane for pure interaura; time shifts (interaural phase delay independant of frequency). Depending on signal and methodology, values have been reported between 2-62 uS, with broadband noise leading to the highest values.
In other words, interaural phase shifts get registred at very short delay times. When you translate this to head movements, you might overlook the importance of this, because they are quite large as compared to the distance time sound travels during 10 uS. But don't forget, both ears travel at the same speed with lateral movements of the head. So, from an engineering standpoint, it is easy to understand how the brain could compensate for this. Therefore the stereo image stays in place when you move your head laterally.
A better measurement for the sensitivity of the auditory system to binaural time delays is to turn your head. As you will have experienced, the stereo image will shift in relation to the direction your nose is pointing, even when you turn your head very slightly. That is the true measure of the sensitivity of the two ears to time shift.
So, when in a discussion of let's say DAC's a claim is made that some set-ups might lead to shifting interchannel time delays or other phase related issues (as I understood Thorsten to do), my attention is sharpened. Btw, after having followed this thread wit interest, am I totally not convinced that this is a real issue for digital, btw.
vac
Thanks for the cite and the thoughts- I'll look for those pages. Did they use speakers or headphones in those experiments?
The interchannel time delay issue was au courant in 1983- shortly thereafter, dual DACs got cheaper and the "problem" disappeared. Doesn't change my basic premise that this is an effect swamped by speaker, seat, and head position, and not one that actually is observed with anything but the oldest and cheapest digital players.
The interchannel time delay issue was au courant in 1983- shortly thereafter, dual DACs got cheaper and the "problem" disappeared. Doesn't change my basic premise that this is an effect swamped by speaker, seat, and head position, and not one that actually is observed with anything but the oldest and cheapest digital players.
Sy, agree, see my last point. But the issue would not seem to be so much static interchannel time delay, but dynamic delay (fluctuating in time) between the two channels. For DAC's I don't see the mechanism through which this could happen, so I don't think it is an issue there.
But for other parts of the chain it might be well overlooked. You mention loudspeakers; passive xovers are likely to be a major source in and by themselves, certainly in combination with the reactive load they offer to an amp. I'd like to learn more from the amp-incrowd to what extent interchannel phase shifts measurements under realistic input and output conditions are part of the measurement suite for advanced amplifier design.
vac
But for other parts of the chain it might be well overlooked. You mention loudspeakers; passive xovers are likely to be a major source in and by themselves, certainly in combination with the reactive load they offer to an amp. I'd like to learn more from the amp-incrowd to what extent interchannel phase shifts measurements under realistic input and output conditions are part of the measurement suite for advanced amplifier design.
vac
Digital Recording
Thanks for all the work! I appreciate it! And I agree with your summary...I know the original discussion was to dither or not to dither, and I was hoping that the fact that it is dither which makes 16 bit usable in a hifi context (especially for acoustic or classical recording) would emerge from the discussion. The fact has been obvious to me since early experiments in the mid-80s. Some of the early Apogee A>D units had great switchable dither spectra which was very educational...
Cheers!
Howie
Howarad Hoyt
CE - WXYC-FM
UNC Chapel Hill, NC
www.wxyc.org
1st on the internet
Thanks for all the work! I appreciate it! And I agree with your summary...I know the original discussion was to dither or not to dither, and I was hoping that the fact that it is dither which makes 16 bit usable in a hifi context (especially for acoustic or classical recording) would emerge from the discussion. The fact has been obvious to me since early experiments in the mid-80s. Some of the early Apogee A>D units had great switchable dither spectra which was very educational...
Cheers!
Howie
Howarad Hoyt
CE - WXYC-FM
UNC Chapel Hill, NC
www.wxyc.org
1st on the internet
Hi,
Yes. The waveforms are all horrible, incidentally and do not match what I expect from REAL Non-Os ADC or DAC.
BTW, how about you show instead the simulation of a real system, instead of making up stuff that we still have not even a feasibility analysis for. I mean someone may mistake what you posted for an actual illustration of what actual dither does in an actual 16 Bit System...
Yes, it matters to my point, if you still use FFT as you do, in the context you have completely missed my point...
Precisely. So you have missed my point.
Do the same as before, take the "ideal" undithered 16 Bit Waveform and the make a version that uses an "ideal" 10 Bit DAC and dither for the rest. Directly overlay the waveforms, it is quite obvious.
It is actually already quite obvious in your pictures...
Ciao T
Yes. The waveforms are all horrible, incidentally and do not match what I expect from REAL Non-Os ADC or DAC.
BTW, how about you show instead the simulation of a real system, instead of making up stuff that we still have not even a feasibility analysis for. I mean someone may mistake what you posted for an actual illustration of what actual dither does in an actual 16 Bit System...
Yes, it matters to my point, if you still use FFT as you do, in the context you have completely missed my point...
Precisely. So you have missed my point.
Do the same as before, take the "ideal" undithered 16 Bit Waveform and the make a version that uses an "ideal" 10 Bit DAC and dither for the rest. Directly overlay the waveforms, it is quite obvious.
It is actually already quite obvious in your pictures...
Ciao T
I think the only way to settle this is to take an ideal sine wave, digitize it, and then feed it into a 16bit NOS DAC. Then take the output and subtract it from the original. Do the same again with a bitstream converter, using exactly the same digitized sinewave.
Then compare the two error waveforms, noise performance etc.
Then compare the two error waveforms, noise performance etc.
Thanks for that link, Scott.
... but ... log3(60dB)? (from page 7)
To borrow an acronym from Thorsten: WTF?
We regularly see log10() for dB ratios, log2() for efficiency with binary processors, and the natural ln(), which is log2.718(), but I've never seen anyone attempt to make practical use of log3() before this paper. 20*log10(2^10) == 60 dB, showing that you need 60 dB attenuation for the AAF before a 10-bit A/D, but how does that lead to log3(60dB)?
Also, I calculate log3(60) as 3.73, not 6.28
What am I missing here?
Any compander is going to have a particular attack time and release time, plus some sort of RMS or other level detection. Each of these elements is going to have some amount of slop, or error. I doubt that any compander will be lossless, but it might be non-objectionable to the listener. Given the cost of 24-bit when DBX introduced their 'solution,' I'm sure that the cost of 24-bit equipment was more objectionable than the losses due to their 16-bit companded option.
To answer your question, I'm sure you could recover something the input audio, but it would sound like it had gone through stages of heavy compression followed by heavy expansion.
Hi,
No need to use real DAC's, all this can be done using "idealised" converters and in something like Mathlab. I lack the time to set this up.
Ciao T
No need to use real DAC's, all this can be done using "idealised" converters and in something like Mathlab. I lack the time to set this up.
Ciao T
For TPDF, the answer is 2 LSB. Other PDFs have different answers. Someone already replied with an answer for RPDF.Trying to put the topic behind me, I guess the piece of information I am missing here is: how much dither is necessary to remove quantizing error as a function of bit depth.
I can't find a definite answer in any of my references, so if someone could shoot me a quick answer to this, I'll stop ragging on the subject
I recommend http://www.ece.rochester.edu/course...ries/2009/1/15_Week_1_files/Lipshitz_1992.pdf
Yes. The second reason why not only "static" interchannel time delay is that ear is most sensitive to rising transients (rising edge of transient signals), rather than steady state. For this reason, the time response of 44.1kHz sampling might be questionable, IMO.
You refer to images, plural, and signals, plural; then later refer to sources. My impression is that we started on the topic of stereo DAC, then it was suggested that a 10 us delay could be heard. It seems that most of the replies have been in regard to stereo, such that when you delay the right channel with respect to the left, there are no "other" signals. The right channel is the only other signal compared to the reference, as in singular, not plural.
If you want to talk about multi-mono mixing consoles, pan laws, binaural recording, multi-microphone techniques, and surround sound reproduction, then I think you're opening an entirely unrelated set of topics. Have mercy on this thread!
As SY seems to have pointed out, nobody has shown any evidence that either (A) there is any perceived difference between a 10 uS delay in the right channel compared to simply moving the right speaker 3.4 mm away from the listener, or (B) that any common CD player or other digital playback system even has a time delay between channels.
By the way, someone pointed out that 2 DACs can be used to time align left and right channels. However, another well-documented technique is to add an analog Sample and Hold circuit to one channel so that the output of 1 DAC can be time-aligned between 2 channels. In systems which already employ a S/H per channel for ZOH, the above can be achieved by adding a 3rd for time alignment.
So, you're claiming that in order to get 3.4 mm closer to the left speaker, you need to move 6"?
Have you assumed an arbitrary constraint that your butt has to stay firmly planted at the rear of your couch? ... or is it acceptable to lean forward ever so slightly as you shift towards the left?
By my trigonometry, if you want to introduce a 10 us time delay between speakers by moving your head, then you only have to move 3.4 mm. All you have to do is constrain your movement to an arc whose radius has the right speaker as the center. This does require you to lean forward, though, but by less than 3.4 mm, so I hope it's not too uncomfortable.
I suppose that if you were to draw all of the intersecting arcs, the actual distance moved might be slightly more than 3.4 mm, but not by much.
Last edited:
Sorry I'm done with this. As John points out the concept of subtractive dither is over 50yr. old in image coding etc. I'm not wasting any more time with the hermitic attitudes of (some) of the audio engineering community. As jcx I find it hard to believe this was not discussed at the Redbook standards meetings. Folks here are refered to the large body of literature and can make up their own minds.
Physically implementing this would be a waste of time, back to the DBT circus.
Well, personally, I'll stick with the bitstream + Digital filter approach I think - my next project.
"Physically implementing this would be a waste of time, back to the DBT circus"
Yes, you are probably right, but at this point the protaganists dont even seem to be able to agree on the hard facts of the two D-A converter topologies.
"Physically implementing this would be a waste of time, back to the DBT circus"
Yes, you are probably right, but at this point the protaganists dont even seem to be able to agree on the hard facts of the two D-A converter topologies.
In fact it is even a bit higher than that, as the "official equation" has an additional term to reflect the ratio between sample rate and target bandwidth.
6.02N + 10log(fsample/2xBW) + 1.76 dB
So, 44.1kHz sample rate and 20kHz bandwidth for red book CD adds another 0.42dB (further oversampling gives of course higher improvement)
@ rsaudio,
somewhat idealized mathematical concepts are fine but we are in fact listening to real world implementations, so i think it is not that surprising that still diverging listening impressions exist, given that interindividual sensory events based on the same acoustic events can be quite different.
It is difficult to seperate the various effects in these real world listening tests, because they all might work together in unpredictable ways.
See for example the NOS/ZOH D/A convertor- it has amplitude effects (due to ZOH), it has variable signal spectral contents effects (due to the low damping effect of the overlying weighting function around the first mirror image), it has different transient behaviour due to missing additional brickwall filtering/rounding/dithering.....
@ SY,
as cited before, Paul Frindle reported in his Convention paper ABX-Results for an impressive list of small differences (done in studios) including a constant halfsample width delay of one channel (leads to ~11µs) and confirmed it later several times in discussions over at the pgm list.
Paul Frindle, Are we Measuring the right Things? Artefact audibility versus measurement
The Measure of Audio, AES UK Conference 1997.
I have to dig deeper in my archive (which suffers in the moment from similar problems as jneutrons), but afair Theile and Wittek should have done some experiments with ITDs in stereophonic presentations, which confirm at least the ~10µs .
As stated earlier, small head movements actually increase the localization accuracy, investigated a couple of times with good accordance.
BTW, this was an interesting description of a dither phenomenon in last years discussion:
http://www.gearslutz.com/board/5985683-post144.html
later on Frindle mentioned differences detectable as low as 4µs.
Last edited:
All the discussion about dither being evil is absurd.
Dither is absolutely necessary to avoid distortion that would appear if high bit depth files would be simply truncated to 16bit.
It is very simple to experiment by generating a 24bit sine wave in some proaudio software (Wavelab) and truncate it to 16bits and also dither it to 16bits. Just do an analysis and also listen. You can draw your conclusions.
For music the truncated version will sound distorted and all the low level information like space cues and reverb tails will be lost while the dithered version will retain all the low level information and no distortion.
Being a mastering engineer and having some highend AD/DA converters and processors I have the luxury to check in real time all these things (real instant A/B test). One of my mastering processors have the ability to check at a push of a button the difference between 24bit truncated and dithered to 16bit. The difference is there even on compressed/loud pop/rock music but on minimalist two microphone recorded classical music or jazz it is a significant difference. Non dithered, it sounds edgy (fuzzy distortion?) and low level info is missing. Dither engaged the fuzzy sound disappear and low level info is there.
The only place where dither is not absolutely necessary is when word length reduction is applied at the final stage from 32bit (or 64bit) floating point or 48bit fixed point to 24 bit final audio file. But if we talk in the context of a digital console or DAW where we have multiple processes on multiple channels, dither is absolutely necessary even at intermediate processes (because errors multiply in the large context).
For better understanding dither read the following documents.
The first one is written in a very intuitive way so even non technical people can understand the essence of dither:
http://www.users.qwest.net/~volt42/cadenzarecording/DitherExplained.pdf
The second is written by Dan Lavry, the guy who has a deep understanding of AD/DA and digital in general:
http://www.lavryengineering.com/white_papers/dnf.pdf
http://www.lavryengineering.com/white_papers/dither.pdf
chrissugar
Dither is absolutely necessary to avoid distortion that would appear if high bit depth files would be simply truncated to 16bit.
It is very simple to experiment by generating a 24bit sine wave in some proaudio software (Wavelab) and truncate it to 16bits and also dither it to 16bits. Just do an analysis and also listen. You can draw your conclusions.
For music the truncated version will sound distorted and all the low level information like space cues and reverb tails will be lost while the dithered version will retain all the low level information and no distortion.
Being a mastering engineer and having some highend AD/DA converters and processors I have the luxury to check in real time all these things (real instant A/B test). One of my mastering processors have the ability to check at a push of a button the difference between 24bit truncated and dithered to 16bit. The difference is there even on compressed/loud pop/rock music but on minimalist two microphone recorded classical music or jazz it is a significant difference. Non dithered, it sounds edgy (fuzzy distortion?) and low level info is missing. Dither engaged the fuzzy sound disappear and low level info is there.
The only place where dither is not absolutely necessary is when word length reduction is applied at the final stage from 32bit (or 64bit) floating point or 48bit fixed point to 24 bit final audio file. But if we talk in the context of a digital console or DAW where we have multiple processes on multiple channels, dither is absolutely necessary even at intermediate processes (because errors multiply in the large context).
For better understanding dither read the following documents.
The first one is written in a very intuitive way so even non technical people can understand the essence of dither:
http://www.users.qwest.net/~volt42/cadenzarecording/DitherExplained.pdf
The second is written by Dan Lavry, the guy who has a deep understanding of AD/DA and digital in general:
http://www.lavryengineering.com/white_papers/dnf.pdf
http://www.lavryengineering.com/white_papers/dither.pdf
chrissugar
I´m sure that Bennett already mentioned the benefit of noise regarding any distortion components during quantizing in his 194x papers, but i have to reread. It happens all the time in technical history that some concepts have to be invented several times to get really known.
BTW, afair Gray and Stockham mentioned that triangular dither was used during the recording of Fleetwood Macs ´Tusk´ in 1979/1980 but was not revealed due to commercial reasons; i think John has already told something about their use of a digital recorder?!
But wrt to dither theory, i think that Widrow and colleagues today argue that, despite theoretical conflicts, gaussian dither is preferable during A/D conversion compared to triangular dither.
As stated earlier, we should be able to agree on the positive effects of dither in various applications, but should be able to accept that real world implementations quite often differ from perfection.
Today most parts of the professional research communitiy agree that perceptual evaluation (by human listeners) of highly processed material is the gold standard , because measurements might not be sufficient.
See for example various ITU-Recommendations like BS.1387 .
- Status
- Not open for further replies.