R/E/P Community

Please login or register.

Login with username, password and session length
Advanced search  

Pages: [1]   Go Down

Author Topic: Are 24 bit converters dithered?  (Read 2203 times)

Jon Hodgson

  • Hero Member
  • *****
  • Offline Offline
  • Posts: 1854
Are 24 bit converters dithered?
« on: July 22, 2009, 07:47:13 AM »

I'm hoping Bruno can shed a little light on this.

Firstly I know that ideally an audio ADC should be dithered, and this should ideally be with something like TPDF noise.

However if we're talking about a 24 bit converter (non oversampling), with its tiny quantization steps then I can't see how you could actually generate that TPDF noise without also generating far more gaussian noise at the same time.

With sigma delta converters the quantization steps are larger, so it seems more doable, though I can't remember seeing any mention of analog dither on converter datasheets (may have missed it, wasn't looking for it at the time), and it would have to be either part of the IC or an input pin, since it would have to go between the integrator and the quantizer within the sigma-delta modulator.

Of course when we're looking at 24 bit audio converters the input signal is invariably going to have several bits worth of gaussian noise on it, making it large and chaotic with respect to a single quantization interval, thus resulting in uncorrelated errors, i.e. noise like in character, though this is not the ideal input noise distribution.

So my question is, are any of these converters dithering at the initial quantizer stage (as opposed to the decimation filters), and if so, how? And if not, how much of a problem is it considering the expected characteristics of the input?
Logged

John Roberts {JR}

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 266
Re: Are 24 bit converters dithered?
« Reply #1 on: July 22, 2009, 07:07:52 PM »

Bruno can give you a better overview but something to chew on in the meanwhile, look at the analog noise floor from the data sheet, and compare that to the LSB quantization level (approximately -6dB per bit).

Any analog noise referenced to the input present during the conversion process will serve to dither the bottom bit(s). The only issue is where exactly in the process does the analog noise floor come in. My guess it is more input related than output.

JR
Logged

Jon Hodgson

  • Hero Member
  • *****
  • Offline Offline
  • Posts: 1854
Re: Are 24 bit converters dithered?
« Reply #2 on: July 22, 2009, 08:07:10 PM »

John Roberts  {JR} wrote on Thu, 23 July 2009 00:07

Bruno can give you a better overview but something to chew on in the meanwhile, look at the analog noise floor from the data sheet, and compare that to the LSB quantization level (approximately -6dB per bit).

Any analog noise referenced to the input present during the conversion process will serve to dither the bottom bit(s). The only issue is where exactly in the process does the analog noise floor come in. My guess it is more input related than output.

JR


Thanks. I'm quite aware of the above though (note the bit in my original post when I talk about the input signal having several bits worth of gaussian noise on it). I also gather from Dan Lavry's comments that his converters at least have no additional dithering at the input.

However ever since someone pointed out on a forum that gaussian noise was not ideal for dithering it's become something of a religious mantra with some people that "you can't rely on self dither", so I'm interested in knowing what the real world situation is to get some perspective on it all.
Logged

John Roberts {JR}

  • Sr. Member
  • ****
  • Offline Offline
  • Posts: 266
Re: Are 24 bit converters dithered?
« Reply #3 on: July 23, 2009, 12:39:48 AM »

I'll leave debates about phenomenon down over 100 dB to others.

JR
Logged

bruno putzeys

  • Hero Member
  • *****
  • Offline Offline
  • Posts: 1078
Re: Are 24 bit converters dithered?
« Reply #4 on: July 23, 2009, 06:00:03 AM »

A/D converters are exceedingly rarely dithered, DAC's usually are.

There are two aspects of dithering in ADC's.

One is simply preventing the quantization of the output word from causing distortion. Theoretically, only dither with at least one RPDF component will do as a true dither source. However, the statistical smoothing of the quantization staircase that results from having a large gaussian noise source is going to be quite sufficient. By "large" I would say anything with a sigma of several LSB's. A typical high-end 24-bit chip has something like 120dB dynamic range, which means that the analogue noise is some 24dB (16 times) higher than the quantisation noise. That is rather enough. It will only start getting mildly problematic as converters pass the 130dB mark.
Note that even though a converter can get by without dithering, that does not mean self-dithering is a valid choice anywhere else. The reason why we need to dither after signal processing even when the signal itself contains a lot of noise is because after processing that noise can no longer be counted on to have a continuous distribution. So, the engineering choice to forgo dither and count on self-dithering has to be made on knowledge of exactly what that noise looks like. This is only the case for A/D converters.

The other aspect of dither is the prevention of tones (whistles, space monkeys, birdies) in sigma delta modulators. Low-bit converters (three or less) can't practically be perfectly (=tpdf) dithered because that would use up all the headroom. What you can do is inject just enough noise in the loop to randomize the tones somewhat. This does not linearize the converter but it does get rid of tones. Since this requires extra hardware people always try to worm out of it, with varying results.
Some designers count on a high-order loop to generate enough noise on its own as to render the likelihood of audible tones small. Common though it may be, it's not an effective strategy. An analogue sigmadelta converter that has no actual measures against whistling will whistle when fed a correctly chosen DC input.
A somewhat effective tactic is to make the noise shaper chaotic, making sustained tones impossible. Unfortunately chaotic noise shapers also produce more outband noise and what you'll find is that the noise increase is more than that of an equally effective added noise source. I did find that a slightly chaotic loop combined with reduced external dither noise can be a tiny bit more optimal than either trick alone but it's hardly worth it.
Probably the best non-dithering tactic is to put so much loop gain in the noise shaper that the quantization noise stays well below the thermal noise floor. That way any tones would simply be masked (but not dithered!!!) by the thermal noise.

Logged
Warp Drive. Tractor Beam. Room Correction. Whatever.

Affiliations: Hypex, Grimm Audio.

Jon Hodgson

  • Hero Member
  • *****
  • Offline Offline
  • Posts: 1854
Re: Are 24 bit converters dithered?
« Reply #5 on: July 23, 2009, 06:23:42 AM »

John Roberts  {JR} wrote on Thu, 23 July 2009 05:39

I'll leave debates about phenomenon down over 100 dB to others.

JR

<sarcasm>
But didn't you know? When somebody's recording of their 16kHz bandwidth microphone through their 90dB snr preamp doesn't sound "right", the obvious thing to obsess about is signal components at 100kHz and 140dB down.
</sarcasm>


Logged

Jon Hodgson

  • Hero Member
  • *****
  • Offline Offline
  • Posts: 1854
Re: Are 24 bit converters dithered?
« Reply #6 on: July 23, 2009, 06:35:26 AM »

Thanks Bruno,

In the second part of your post where you're talking about sigma delta modulators, are you talking about the implementations in ADCs or DACs or both?

I've been thinking about implementing a low order sigma delta DAC on an FPGA (part of an idea for a project, perfect fidelity would not be necessary, in fact a certain amount of "character" would be fine), so that part is particularly interesting to me (if rather daunting).
Logged

bruno putzeys

  • Hero Member
  • *****
  • Offline Offline
  • Posts: 1078
Re: Are 24 bit converters dithered?
« Reply #7 on: July 23, 2009, 08:23:48 AM »

Analogue and digital sigmadelta modulators (for use in ADC's and DAC's respectively) have the same issues regarding tones and limit cycles, and the methods used to address them are the same as well. It's important to stress that because in one paper even John Vanderkooy surmised that effects like bilateral symmetry of the spectrum found in 1-bit converters were reduced in analogue modulators because of electronic noise. So I tried it and found that most ADC's simply had enough DC offset to pull the modulator away from this mode, but that if you trimmed it to close to zero even a continuous time modulator will cheerfully produce a mirror-image spectrum.

Analogue modulators have some extra issues (like whistling when there is crosstalk between the data and the clock) but managing these by adding noise inside the loop is not very productive.

There's one thing where the self-dither crowd is somewhat correct, and that's the fact that a low-order modulator is more generous with whistles. So unless you have technical reasons for keeping the order low I'd go for at least a 5th order design.
Logged
Warp Drive. Tractor Beam. Room Correction. Whatever.

Affiliations: Hypex, Grimm Audio.
Pages: [1]   Go Up
 

Site Hosted By Ashdown Technologies, Inc.

Page created in 0.051 seconds with 22 queries.