Hi,
As a different datapoint.
Following a similar debate in the 90's a famous UK designer (who was adamant on remaining incognito) measured his creation which combined the then latest Cirrus Logic DAC with 5532 and TL082 Op-Amp's.
I suggested noise load measurements and he was able to make his AP jump though this particular hoop. His measurements at digital full scale, on the analogue output after his analogue stage (which being a steep lowpass obviously acts to apply massive averaging) showed a noisefloor modulation of -40dB, or 1% Fuzzy distortion AFTER filtering and averaging.
Now this is a different type of noise modulation than that of noisefloor in the presence of DC and may have been in part TIM from his op-amp's unsuited to deal with the massive amounts of supersonic noise the DAC put out.
Ciao T
As a different datapoint.
Following a similar debate in the 90's a famous UK designer (who was adamant on remaining incognito) measured his creation which combined the then latest Cirrus Logic DAC with 5532 and TL082 Op-Amp's.
I suggested noise load measurements and he was able to make his AP jump though this particular hoop. His measurements at digital full scale, on the analogue output after his analogue stage (which being a steep lowpass obviously acts to apply massive averaging) showed a noisefloor modulation of -40dB, or 1% Fuzzy distortion AFTER filtering and averaging.
Now this is a different type of noise modulation than that of noisefloor in the presence of DC and may have been in part TIM from his op-amp's unsuited to deal with the massive amounts of supersonic noise the DAC put out.
Ciao T
Hi,
PRECISELY.
Dither actually reduces the ability to distinguish between two quantities and thus lower resolution.
It also increases the difference between a measured value and the true value in a sufficiently large dataset (standard deviation) and thus lowers accuracy.
As it also dramatically increases the error band it also reduces the repeatability and thus precision.
Or in other words if you apply what I like to call "changing goalposts", "hand waving", "magic" and "the power of faith". It actually is the opposite of what happens in REALITY.
Ciao T
PRECISELY.
Dither actually reduces the ability to distinguish between two quantities and thus lower resolution.
It also increases the difference between a measured value and the true value in a sufficiently large dataset (standard deviation) and thus lowers accuracy.
As it also dramatically increases the error band it also reduces the repeatability and thus precision.
Or in other words if you apply what I like to call "changing goalposts", "hand waving", "magic" and "the power of faith". It actually is the opposite of what happens in REALITY.
Ciao T
Dither
My experience recording digital is that reverb tails and other low-level audio prematurely distort at low levels when no dither is employed. The spectral shaping of the dither can make a real difference to the overall noise floor and distortion profile at low levels.
So to summarize my feelings on dither:
Analog's strength is high resolution, but with low accuracy (s/n+d).
Digital's strength is high accuracy (when one employs many bits), but low accuracy AND resolution (at low levels) .
The fewer bits one uses, the poorer the resolution and accuracy. When a signal is so low it only utilizes a couple of bits it is stored in a very low resolution and highly inaccurate way.
In order to make digital storage better we utilize the high accuracy of the upper bits, and add one of the weakest aspects of analog (noise) to the lowest 2 bits to increase resolution. I would guess that the accuracy (s/n+d) of this system at this point is probably equal to or worse than an analog system with similar weighted s/n.
So dithered digital systems should be called digilog...wait, that's been used...anatal...no, that sounds like a hemorrhoid medicine...we'll think of something else.
It reminds me that no matter how digital a system gets, it is often terribly analog in it's operation...like optical discs, which are analog storage media no matter how you slice it (pun intended !)
Howie
Howard Hoyt
CE - WXYC-FM
UNC Chapel Hill
www.wxyc.org
1st on the internet
My experience recording digital is that reverb tails and other low-level audio prematurely distort at low levels when no dither is employed. The spectral shaping of the dither can make a real difference to the overall noise floor and distortion profile at low levels.
So to summarize my feelings on dither:
Analog's strength is high resolution, but with low accuracy (s/n+d).
Digital's strength is high accuracy (when one employs many bits), but low accuracy AND resolution (at low levels) .
The fewer bits one uses, the poorer the resolution and accuracy. When a signal is so low it only utilizes a couple of bits it is stored in a very low resolution and highly inaccurate way.
In order to make digital storage better we utilize the high accuracy of the upper bits, and add one of the weakest aspects of analog (noise) to the lowest 2 bits to increase resolution. I would guess that the accuracy (s/n+d) of this system at this point is probably equal to or worse than an analog system with similar weighted s/n.
So dithered digital systems should be called digilog...wait, that's been used...anatal...no, that sounds like a hemorrhoid medicine...we'll think of something else.
It reminds me that no matter how digital a system gets, it is often terribly analog in it's operation...like optical discs, which are analog storage media no matter how you slice it (pun intended !)
Howie
Howard Hoyt
CE - WXYC-FM
UNC Chapel Hill
www.wxyc.org
1st on the internet
people have listened to dither, in addition to theorizing
Wannamaker is surprisingly readable for a PhD thesis -
http://www.digitalsignallabs.com/phd.pdf
but for the lazy:
Wannamaker is surprisingly readable for a PhD thesis -
http://www.digitalsignallabs.com/phd.pdf
but for the lazy:
I'm totally unclear how you've arrived at those numbers. My TDA1543 does considerably better than 30% THD+N at -60dBfs and its an 'economy class' part.
Reality is an illusion, albeit a particularly persistent one. Care to say what form of 'reality' you're referring to? Or alternatively what measurements show there's 6% fuzzy distortion within the audio band at 0dBfs?
Haven't been arguing about what's 'perceived as neutral/beneficial'. I'd myself state that the process of quantization is what adds 'fuzzy distortion' at ANY level, dither is an attempt to mitigate that. How well it actually does what's claimed for it I haven't reached clarity on personally, but I'm heading in that direction.
Hi,
What is the peak level of the dither? What is the peak level of the signal? What is the difference rendered as %?
It will have that level of "fuzzy" distortion (as opposed to harmonic distortion) with dither.
BTW, I first played with a kusonoki style DAC (CS8412 & TDA1543, no oversampling, output via 1:1 pro audio transformers) in the late 90's, just like you it convinced me that there was something wrong with the low bit converters...
Only up to a point. There are certain unalterable, physical realities which are independent from the mind that perceives them.
Ciao T
What is the peak level of the dither? What is the peak level of the signal? What is the difference rendered as %?
It will have that level of "fuzzy" distortion (as opposed to harmonic distortion) with dither.
BTW, I first played with a kusonoki style DAC (CS8412 & TDA1543, no oversampling, output via 1:1 pro audio transformers) in the late 90's, just like you it convinced me that there was something wrong with the low bit converters...
Only up to a point. There are certain unalterable, physical realities which are independent from the mind that perceives them.
Ciao T
Thanks for this link.
By the way, one of the criticisms of the Vanderkooy, et al, papers is that what they claim is not true just because they claim it is true. On that point, I will say that I do not personally take every claim they make as the whole truth without limit. For example, the claim that TPDF is a minimum, and that nobody could tell the difference beyond that level, is something I've seen in a number of papers. There are two ways to respond to this. On the positive side, I will agree that TPDF is an absolute minimum. It's fairly easy to prove that 1st order or 0th order distribution functions produce inadequate dither that still allows quantization distortion in the PCM. On the negative side, though, I would not accept those tests as the final word. Perhaps, some day, with better circuits and better tests, we may learn that higher orders beyond TPDF provide even better results. In that respect, I do not agree 100% with the papers that would stop at 2nd order, even though I agree that less is certainly insufficient.
I believe it was Thorsten who suggested that comparing TPDF dither to Gaussian Johnson noise is an unfair comparison. On that particular point, I would agree, at least in theory. I'd love to hear the results of experimenting with Gaussian dither, even though that would require transcendental computations to produce in the digital realm (or the dither could be produced in the analog realm).
As for the other claims that dither is some monster with the special power to destroy audio, I just don't buy it. I see no difference between dither noise and other sources of noise. Resistor noise and tape noise do not completely destroy the audio, even though it's best to minimize them. Dither noise should be the same, unless the characteristics of some particular dither noise source are pathological.
Also, Thorsten keeps going on and on about massive amounts of dither. None of the papers I've cited have recommended anything beyond 2 LSB of dither, so I don't see the point of this particular criticism - it's non sequitur. Granted, 1-bit (sigma-delta) converters do employ varying degrees of dither in excess of 2 LSB, but none of the papers I've cited speak favorably of that approach to dither. In other words, I think 'dither' is perhaps too broad a category in which to lump every one of the techniques.
If we're talking 16 bits then its 2LSB, TPDF that means about -90dBfs.
0dBfs, about 0.003% I make it.
Well I don't (so far) dither DACs, just send them the bits I find on the CD. Thus the 'fuzzy distortion' must be reflected in the figures NXP/Philips give in their datasheets right? After all its not 'THD', its 'THD+N'.
Ah my experience goes back a bit earlier than yours, prior to the days of the CS receivers I was fiddling around with YM3623B or something such like, and AD1862.
And you know this how?
Hi,
And then:
WTF? Resistor and tape noise is gaussian, dither is not, so I see a difference right away...
Listen to an AAD processed CD.
The Tape noise is gaussian AND pink noise in nature (the opposite of commonly applied dither) and acts a special kind of dither. I believe one or two "mastering plugins" offer such "tapenoiselikedither" as an option, others prefer to sample rate and wordlenght downconvert their recordings to 16 Bit by interposing an analogue tape machine, I know two quite famous people who use (or at least used to use) a Studer G36 (modified) for that job...
Okay, it may be that no-one speaks favourable of these low bit DS converters with their massive levels of dither and noise shaping and dither (though certainly Lipshitz/Vanderkoy do so repeatedly), however these converters are now the absolute industry standard for recording and playback, despite this.
It is in recognition of the these facts that I criticise such use of dither.
We may argue about dither (especially when downconverting HD recordings to 16 Bit), but I have no major problem with people using it or not.
It is a rather different story if we see a 128 Times oversampling (@ single speed - less at higher speed) converter with a 25 Level modulator and then see people lyrically waxing how this latest 32 Bit DAC (which at 192KHz sample rate is actually a 9.5 Bit DAC with 11.5 Bit worth of dither) lauded as the solution that finally brings new levels of resolution to audio.
Ciao T
And then:
WTF? Resistor and tape noise is gaussian, dither is not, so I see a difference right away...
Listen to an AAD processed CD.
The Tape noise is gaussian AND pink noise in nature (the opposite of commonly applied dither) and acts a special kind of dither. I believe one or two "mastering plugins" offer such "tapenoiselikedither" as an option, others prefer to sample rate and wordlenght downconvert their recordings to 16 Bit by interposing an analogue tape machine, I know two quite famous people who use (or at least used to use) a Studer G36 (modified) for that job...
Okay, it may be that no-one speaks favourable of these low bit DS converters with their massive levels of dither and noise shaping and dither (though certainly Lipshitz/Vanderkoy do so repeatedly), however these converters are now the absolute industry standard for recording and playback, despite this.
It is in recognition of the these facts that I criticise such use of dither.
We may argue about dither (especially when downconverting HD recordings to 16 Bit), but I have no major problem with people using it or not.
It is a rather different story if we see a 128 Times oversampling (@ single speed - less at higher speed) converter with a 25 Level modulator and then see people lyrically waxing how this latest 32 Bit DAC (which at 192KHz sample rate is actually a 9.5 Bit DAC with 11.5 Bit worth of dither) lauded as the solution that finally brings new levels of resolution to audio.
Ciao T
Hi,
I asked Peak. Make that -84dB...
But I did slip somewhere in the calculations.
So at -60dBfs it is only 6% fuzzy distortion, not 30%...
Which probably contains enough noise from the recoding chain to have it's own dither, plus you forget the ultrasonic images from a lack of digital filtering which may act as signal dependent ultrasonic dither (I believe this was proposed by Douglas Rife from DRA Labs, but I may remember wrong)...
Nope, it is not, as Philips did not use dither.
BTW, the measurable effect of dither is quite clear and easy to see on any AP2, as you can select different forms and levels of dither or non at all. If I want to know what is going I ALWAYS measure un-dithered, with dither the measurements present incorrect results, especially stuff like J-Test and FFT's.
I fiddeled with a lot of stuff before this Kusonoki DAC (including the Yamaha receiver, as well as the Philips one), in fact, I was absolutely convinced that it could not work and would be VERY BAD, given how my much more advanced DAC's sounded so poor...
Empirically, every time I look at my bank accounts balance.
Ciao T
I asked Peak. Make that -84dB...
But I did slip somewhere in the calculations.
So at -60dBfs it is only 6% fuzzy distortion, not 30%...
Which probably contains enough noise from the recoding chain to have it's own dither, plus you forget the ultrasonic images from a lack of digital filtering which may act as signal dependent ultrasonic dither (I believe this was proposed by Douglas Rife from DRA Labs, but I may remember wrong)...
Nope, it is not, as Philips did not use dither.
BTW, the measurable effect of dither is quite clear and easy to see on any AP2, as you can select different forms and levels of dither or non at all. If I want to know what is going I ALWAYS measure un-dithered, with dither the measurements present incorrect results, especially stuff like J-Test and FFT's.
I fiddeled with a lot of stuff before this Kusonoki DAC (including the Yamaha receiver, as well as the Philips one), in fact, I was absolutely convinced that it could not work and would be VERY BAD, given how my much more advanced DAC's sounded so poor...
Empirically, every time I look at my bank accounts balance.
Ciao T
Perhaps I'm being super-dumb here but I can't quite make the leap from 1LSB peak to -84dBfs. Care to help me out a little - what am I missing?
Still the typicals for the 1543 beat your figure, so something's amiss.
More hand-holding required - please set out how Philips made that measurement for us? I suspect the DNL performance of the 1543 is worse than 2LSB of TPDF dither btw.
Ah the SAA7274? I got samples for it and some BB212 varactors and 11.2896MHz XTALs too but was too lazy to even get around to breadboarding it. I found the datasheet a bit indigestible I seem to recall.
Empirically, every time I look at my bank accounts balance.
Perhaps one or more of Robert Kiyosaki's literary offerings would help you regain perspective?
Hi,
Peak to peak level of the dither is 4.
Peak to peak level of a full scale signal is 65536.
So 65536/4 = 16384 = 84dB
No, nothing amiss.
The TDA1543 was characterised by Philips without dither and hence lacks this particular type of fuzzy distortion. Just measure it on an AP2 using the analog analyser on the output and compare dither on/off.
You mean you cannot read the datasheets by Philips where they describe the method?
They used sinewave coefficients generated via EEPROM's, as was common in those days, no dither, standard analogue stages and suitable analysers (this was the pre-AP2 era, so measurements where actually less fancy in their results but often more meaningful).
Ciao T
Peak to peak level of the dither is 4.
Peak to peak level of a full scale signal is 65536.
So 65536/4 = 16384 = 84dB
No, nothing amiss.
The TDA1543 was characterised by Philips without dither and hence lacks this particular type of fuzzy distortion. Just measure it on an AP2 using the analog analyser on the output and compare dither on/off.
You mean you cannot read the datasheets by Philips where they describe the method?
They used sinewave coefficients generated via EEPROM's, as was common in those days, no dither, standard analogue stages and suitable analysers (this was the pre-AP2 era, so measurements where actually less fancy in their results but often more meaningful).
Ciao T
This seems like a good time to point out one of the tenets of dithering: Any time the bit depth is reduced, you (probably) need dithering. Since analog is equivalent to infinite bit depth, you need dithering during A/D, and the amplitude of the dither is 2 LSB. Similarly, when you reduce from 24-bit to 16-bit, dither is needed, again at 2 LSB (this time, though, the dither is much higher than the 24-bit A/D). Where we start to lose consensus (as if we had it thus far) is when the word length is reduced from 48-bit to 24-bit, or from 80-bit to 40-bit (SHARC).
Anyway, my point is that if a particular DAC implements DSP at some bit depth greater than it's converter, then dithering is probably a good idea. Personally, I doubt that I would prefer a DAC that manipulates the bits this way, but there you have it. That's still not an example where you would manipulate the bits from the CD, but rather that the DAC would be manipulating them for you.
Another example is lossy compression. While I do not prefer MP3 or AAC, there is some modern music which is only available in that format. When listening to lossy sources, I have have my MHLabs MIO FireWire audio interface programmed with DSP blocks that perform crossover EQ and 24-bit TPDF dither. Their driver delivers 32-bit floating point from the MP3 or AAC decoder directly to the DSP over FireWire, and the 32-bit samples are dithered down to the 24-bit DAC outputs rather than using awful truncation. I also reduce the gain digitally to avoid digital clipping with the multiple MP3 frequency bands are mixed after reconstruction from the frequency domain. It's a bit like polishing a turd, but it does seem to make MP3 much more tolerable in most examples.
Where did you get that number? Every reference I recall shows the peak to peak level as 2, not 4. The deviation is 1 LSB, and thus the peak-to-peak is 2 LSB. I suppose there's more than one way to skin a cat...
Ah that's where we differ - 2LSB dither means +/-1LSB, so peak-peak level is 2, not 4.
How do you know Philips didn't dither the test waveform? I don't have an AP2 nor do I wish to acquire one. Did use its predecessor for a while though.
Well I've read the TDA1543 datasheet through a few times and found no mention of lack of dither, nor EEPROMs. Just they tested at 192k sample rate.
Edit that to say: "I see no substantive difference between dither noise and other sources of noise."
For one thing, not all dither is TPDF. For another, I'd say that -60 dB of tape noise is more objectionable than -90 dB of dither noise, regardless of distribution.
Just an aside here, since I've been talking about Philips quite a bit. In my study of the venerable SAA7220 digital filter I found no evidence at all that Philips included dither at the final truncation from their FIR filter accumulation (potentially 28-32bit word length, depending on how they did it). They do though include an offset, in the sample I had it was 0x0020.
Hi,
We have no consensus.
To give you a simple and practical example.
Take an LP, record at 88.2K (or 176.4KHz if you like) and 24 Bit.
Then post process to make best use of 16 Bit dynamic range (eg set highest peaks to 0dBfs possibly even compress some if you feel like it).
Then convert to 44.1KHz/16Bit.
Do one conversion using ZOH & Truncation (essentially discard every 2nd sample and the lower 8 Bit of the 24 Bit signal) and another with the usual ASRC and Dither.
Listen to the resulting files in direct comparison to the original LP.
Note I am aware that the original LP has a noisefloor that is by far higher than that of dithered 16 Bit or indeed 24 Bit Audio, still do try.
So all those low bit audio ADC's (100% of the market) and low bit audio DAC's (99.999% of the market, only one traditional DAC chip remains in production and another has been announced but is not shipping silicone) that use more dither than real bit's are a good thing then?
Then I sincerely hop you are running a multibit DAC without dither added by the digital filter (preferably without digital filter at all)...
I though you did not want a DAC "that manipulates the bits this way", yet what you use has DAC Chips (AKM AK4395) and ADC's chips (AKM AK5394) that I previously described quite accurately as 9.5 Bits plus 11.5 Bits of Dither!
What gives?
Ciao T
BTW, I evaluated both Cirrus Logic's and AKM's "32 Bit DAC's" (their design and technologies are nearly identical) and they did not get design wins in any products I have been involved with. There is worth out there, but why set the standards so low?
We have no consensus.
To give you a simple and practical example.
Take an LP, record at 88.2K (or 176.4KHz if you like) and 24 Bit.
Then post process to make best use of 16 Bit dynamic range (eg set highest peaks to 0dBfs possibly even compress some if you feel like it).
Then convert to 44.1KHz/16Bit.
Do one conversion using ZOH & Truncation (essentially discard every 2nd sample and the lower 8 Bit of the 24 Bit signal) and another with the usual ASRC and Dither.
Listen to the resulting files in direct comparison to the original LP.
Note I am aware that the original LP has a noisefloor that is by far higher than that of dithered 16 Bit or indeed 24 Bit Audio, still do try.
So all those low bit audio ADC's (100% of the market) and low bit audio DAC's (99.999% of the market, only one traditional DAC chip remains in production and another has been announced but is not shipping silicone) that use more dither than real bit's are a good thing then?
Then I sincerely hop you are running a multibit DAC without dither added by the digital filter (preferably without digital filter at all)...
I though you did not want a DAC "that manipulates the bits this way", yet what you use has DAC Chips (AKM AK4395) and ADC's chips (AKM AK5394) that I previously described quite accurately as 9.5 Bits plus 11.5 Bits of Dither!
What gives?
Ciao T
BTW, I evaluated both Cirrus Logic's and AKM's "32 Bit DAC's" (their design and technologies are nearly identical) and they did not get design wins in any products I have been involved with. There is worth out there, but why set the standards so low?
Hi,
1 LSB = 2 Levels (as in using the single LSB)
2 LSB = 4 Levels (as in using the two lower Bits)
If this is not your use then it can be only 1 LSB across several samples, as in each individual sample there in only 1 LSB.
Then the dither occupies 1LSB.
Ciao T
PS, it may be my take is archaic, it goes back to the use we had in industrial systems in the 80's...
1 LSB = 2 Levels (as in using the single LSB)
2 LSB = 4 Levels (as in using the two lower Bits)
If this is not your use then it can be only 1 LSB across several samples, as in each individual sample there in only 1 LSB.
Then the dither occupies 1LSB.
Ciao T
PS, it may be my take is archaic, it goes back to the use we had in industrial systems in the 80's...
- Status
- Not open for further replies.