OK, everybody, please listen up:
This is what I said in an interview for 'TAS' more than 31 years ago. This issue is Mar. 1980. p.58. 'TECHNOCRACY. Frontiers of Thought: The Flaws of Contemporary Digital Recorders, The Problems of Spiegel's Black Box, The Promise of Improved Analog Recorders.'
" The same thing happens with what we call 'dithering' in digital. Dithering is adding a pseudo random noise to the baseband at very low levels in an attempt to mix up the distortions so that they measure not as distortion, but as noise. This is an interesting solution, and it may be a very good solution, but it does not improve the accuracy of the system. The capturing of the signal will always be related to the number of bits being used at that instant. It doesn't matter if you can measure it as a tone, or whether you measure it as noise, or instability or a phase shift. It will always be the number of bits that will determine the overall accuracy."
This is what I said in an interview for 'TAS' more than 31 years ago. This issue is Mar. 1980. p.58. 'TECHNOCRACY. Frontiers of Thought: The Flaws of Contemporary Digital Recorders, The Problems of Spiegel's Black Box, The Promise of Improved Analog Recorders.'
" The same thing happens with what we call 'dithering' in digital. Dithering is adding a pseudo random noise to the baseband at very low levels in an attempt to mix up the distortions so that they measure not as distortion, but as noise. This is an interesting solution, and it may be a very good solution, but it does not improve the accuracy of the system. The capturing of the signal will always be related to the number of bits being used at that instant. It doesn't matter if you can measure it as a tone, or whether you measure it as noise, or instability or a phase shift. It will always be the number of bits that will determine the overall accuracy."
Last edited:
I still don't get it - doesn't this just show that dither is noise? What am I missing?
Wasn't the CBS disk made in a computer? Been a while since I had one, but my famously bad memory is that it also had illegal values.
Thanks,
Chris
OK John lets drop it, your comments don't contain any formulation of a technical discussion that two sides could engage in. There is definately a possibility that at the roll out of CD Blesser's work was a threat to the prevailing hegemony. You were snubbed at the AES as I was by Lipshitz himself, he acted like I had nothing of value to say. In light of this I suggest we just get back to some technical discussion.
Last edited:
Yes, but in the case of 1kHz at -90dB with a simulated ideal A/D there is inband THD (3Khz, 5kHz, ...) that are not insignificant. The simple addition of noise and these do not appear in the spectrum. The remainder is a noise floor that no one has demonstrated as anything but benign. I don't see any signal related corelations, etc.
Measuring the output of a noise-shaping DAC is moving the goal posts. That's because there's dither added within those devices at a much higher level because there's a much, much coarser quantizer inside. I haven't read the datasheet in depth but it looks like that chip adopts 13-level PWM. In such a case I'd be surprised if the dither (now around -20dB in a much wider bandwidth than the original 20kHz) wasn't visible on a scope, decent or otherwise.
I wasn't snubbed by Dr. Lipshitz, we had long discussions in the hallway. I was a 'big shot' in those days, consultant for HK, paper giver at IEEE, Tulsa and Tokyo Audio Show. Used to wear a suit and tie. However, I declined the invitation to be president of the AES as it entailed too much work and I had no support staff.
You should read the rest of my interview, I mentioned Barry Blesser on the same page.
You should read the rest of my interview, I mentioned Barry Blesser on the same page.
"98.08 dB" is not a technical term, it is a numerical quantity.Again, please be correct in technical terms. In theory, it is 98.08 dB (6.02N + 1.76). When you start adding FFT narrow band analysis with narrow binwidth, long memory and averaging, you can get to fantastic 130 - 140dB and confuse the issue
And, technically, it's (6.0206N - 1.76) ... you have to be the first person I've heard suggest that 16-bit CD has over 98 dB of S/N rather than the typically-quoted 96 dB. At the moment, I can't cite a reference, but I've never seen any claim that the S/N or dynamic range is better than the bit depth. This is all referring to maximum signal amplitude versus quantization noise.
I will agree, though, that the resolving ability of 16-bit might extend down to -130 dB or -140 dB with the right front-end processing.
Sorry it was Bob Pease, so there we can't cross domains. You were not complimentary toward Dr. Lipshitz when you used to wake me up at 11PM (couldn't figure out the 3Hr. difference). Peace, we need it.
Last edited:
The only evaluation I've done of software vinyl correction was carried out at 192 kHz. If you can show me a piece of vinyl with frequency content exceeding 96 kHz, which can be resolved by the needle, cartridge, and preamp, then I might buy into your discrediting of the digital variants.
Your description has only served to convince me that I was correct to avoid any attempt to use 44.1 kHz sampling for vinyl, but I already knew that.
I released a free AudioUnit plugin that implements RIAA decoding, and one of the corner frequencies is at 50 kHz! Although the plugin is 'compatible' with lower sampling rates, the RIAA curve is not nearly as accurate at low sampling frequencies.
Agreed. 'magic' was a bad choice of words on my part; I was merely trying to save time and space.
EDIT: Ultrasonic processing is equally possible with digital audio, it merely requires higher sample rates.
Last edited:
I agree Chris its all getting very tangled, hard to separate out the different strands. Here's my take on it, FWIW.
ThorstenL is conflating what he hears in sigma-delta dacs, which I take to be 'fuzzy distortion', with the effects of dither at the 16bit level. I also hear 'fuzzy distortion' in sigma-delta dacs, I equate this with noise modulation possibly caused by having the wrong kind of dither in the noise shaping loop.
Going back to theory of dither, we must have TPDF dither to avoid level-dependent quantization noise. RPDF dither introduces noise modulation, that's well documented. So what I'd like to see is a treatment of how we're sure that the re-quantization involved in the noise-shaping loop is indeed dithered in TPDF form, rather than some other, ill-defined dither probability density function.
Bottom line is - if the probability density function of the dither noise provided to the shaping quantizer isn't triangular (or perhaps gaussian?), we're going to be getting noise modulation. Given this quantizer is relatively coarse, the noise modulation is going to be all the more audible. That's what I take to be 'fuzzy distortion' in the way that ThorstenL is calling it.
Scott - is this a question you can address to Bob Adams?
Hi John,
I'm sure you won't be surprised when I point out that audio is a complicated and multifaceted phenomenon. Everything that you wrote in the quote above is correct, but that does not mean it tells the whole story. You've left out one very important factor: Without dither, signals with an amplitude of less than 1 LSB completely disappear, unless their DC offset happens to place the waveform right at the threshold between two codes. Similarly, small signal amplitudes which cross the code threshold are boosted to 1 LSB in amplitude and simultaneously converted to a square wave. These two effects represent extremely nonlinear performance that is very correlated and thus highly objectionable.
As soon as dither is added to the signal, the PCM can preserve input amplitudes well below 1 LSB. With proper dither, the nonlinear distortion completely disappears. The only cost is noise of 2 LSB or less, and with 24-bit A/D that's less significant than Johnson noise or other common Gaussian sources.
The only reason I moved from the "dither is awful noise/distortion crowd" to the "dither is absolutely necessary" crowd is because I witnessed this first hand in my own, early digital circuits prior to 1984. The phenomenon is much easier to witness with an 8-bit A/D and an adjustable analog DC offset. It's fairly easy to feed a signal to an 8-bit A/D where the signal amplitude is less than 1 LSB. Then, as you vary the DC offset, you can hear the input signal converted into one of two results: Either it becomes massively distorted as a square wave (100% clipping, but 1 LSB in amplitude), or it disappears completely into digital silence. Once you've witnessed this, you will understand why dither is necessary (unless, of course, other noise sources provide sufficient dithering on their own).
In other words, your analysis has completely ignored signal components that are less than 1 LSB in amplitude, and that is why your analysis is not fully correct.
Note that the dither does not have to be "pseudo random" - you can actually design it to be truly random, especially if you're depending upon analog noise sources.
You never gave me your phone number. No *69 back then, I wouldn't have paid for it even if it was available.
Last edited:
'How do you ensure the noise shaping (re)quantizer doesn't introduce noise modulation?'.
How could anyone be woken at 11PM?
In flagrante delicto, perhaps. Or is that too much information?
- Status
- Not open for further replies.