Why do NOS dacs sound different to oversampling designs?

[j_j]

http://www.aes-media.org/sections/pnw/ppt/jj/adc.ppt Once again offered. Sadly, no recording that I know of.

 

[solderdude]

a small question regarding dither.
You say its a requirement, is this also true for 24bit PCM, say 24/96 ?
I would argue the dither would be lost in real world noise.
Technically it will be better but in practice will it have benefits ?

 

[SIY]

Not JJ, so PMFJI, but the "real world noise" actually is dither. This is something we discovered in FT infrared measurements when I was fresh out of school and working at Nicolet- the detector noise was an excellent way to increase apparent resolution.

 

[j_j]

a small question regarding dither.
You say its a requirement, is this also true for 24bit PCM, say 24/96 ?
I would argue the dither would be lost in real world noise.
Technically it will be better but in practice will it have benefits ?

At 24 bits it's a philosophical question, really. Do you want to capture the system noise from before the ADC?

Now, for some things (seismic sensing, etc) there might be an issue, but the noise is still likely to be larger than a few times 1 LSB.

 

[j_j]

Not JJ, so PMFJI, but the "real world noise" actually is dither. This is something we discovered in FT infrared measurements when I was fresh out of school and working at Nicolet- the detector noise was an excellent way to increase apparent resolution.

If you can take a long, long mean in long-word arithmetic, or even use a more appropriately shaped filter, yes.

 

[solderdude]

I thought dither was shaped in spectrum and can only be dither when it is 1LSB in size, otherwise it is just noise even though it works a bit like dither.

edit: ahh.. was crossposting with J_J

 

[j_j]

I thought dither was shaped in spectrum and can only be dither when it is 1LSB in size, otherwise it is just noise even though it works a bit like dither.

edit: ahh.. was crossposting with J_J

Dither does not have to be shaped, white is just fine, but strictly speaking, must have a precise (triangular) PDF and a precise RMS value.

When system noise is 100 times that, the question becomes "do you correctly characterize the system noise". Strictly speaking, you should if you're documenting something. It probably doesn't matter as much when you're recording.

 

[DonH56]

Dither (noise decorrelation) reduces (quasi-)correlated quantization noise at the cost of reduced overall SNR. I used colored dither decades ago in a radar system. Correlated spurs of any sort are very undesirable in a radar system (is it a target or an artifact?) The noise decorrelation used was fairly narrowband (I think it was 10~15 MHz relative to the wideband 1 GHz or so radar signal bandwidth) and restricted to a fairly low-frequency band but well (MHz) above DC as we needed very LF phase information for signal reconstruction (I/Q system). The amplitude ranged from <1 lsb to ~25 lsbs or so; I think I finally settled on 10~15 lsbs as the best for that system. I also played around with different noise sources and waveshapes but honestly don't recall the differences; not sure they were significant, and more importantly needed to use an essentially filtered Gaussian noise source we had available rather than adding another specialized signal generator to the design. SWaP (size, weight, and power) were important design goals (nose cone radar). I also had to control the amplitude carefully (I used a fancy soft saturation circuit). It was a fun project and I liked moving beyond my IC-designer's world up the chain a bit to see how my ADC was used and how I could improve it at the system level.

For delta-sigma designs, again noting it has been a while since I messed with them and they usually targeted RF rather than audio, dither was sometimes (also) injected internally to suppress tones (large spikes due to finite digital filter lengths) in intermediate stages of the converter (ADC or DAC). The level and bandwidth varied with the design. Most if not all audio converters these days are multistage and include multibit converters inside the delta-sigma loop, so tones are less an issue (the more complex the architecture and more stages are added the less likely tones will be a problem), but having a more "white" or "pink" noise floor sounds better than the raw quantization noise (more sawtooth-like so a little harsher-sounding) so dither lives on. There have also been papers presenting noise decorrelation (dither) as helping reduce the impact (sound) of noise modulation in delta-sigma converters.

As JJ (correctly, of course) said dither is critical in very-LF sensing systems where a lot of instrumentation-grade 24-32 bit converters are used. There signal frequencies are very low and periods far exceed the length of the filters in the converters so dither is used to provide a "clean" noise floor. Signals can be extracted from well below the noise floor (our ears and brain can do that, too, natch).

FWIWFM - Don

 

[dc655321]

At 24 bits it's a philosophical question, really. Do you want to capture the system noise from before the ADC?

I have been pondering the question of dithering a 24 bit audio signal too.

Take the scenario of a convolution engine, processing a 16 bit, 44.1kHz input signal with some FIR filter(s).
If the engine works internally at 32-bit floating point by first normalizing its input signal to +/- 1.0, convolves the input and filter(s), then de-normalizes to a 24 (or 32 bit) integer for output (non-truncating rounding method), is there anything to be gained by applying dither to the output?

Me thinks not, but I would be grateful to have my notions reinforced or destroyed

 

[mansr]

I have been pondering the question of dithering a 24 bit audio signal too.

Take the scenario of a convolution engine, processing a 16 bit, 44.1kHz input signal with some FIR filter(s).
If the engine works internally at 32-bit floating point by first normalizing its input signal to +/- 1.0, convolves the input and filter(s), then de-normalizes to a 24 (or 32 bit) integer for output (non-truncating rounding method), is there anything to be gained by applying dither to the output?

Me thinks not, but I would be grateful to have my notions reinforced or destroyed

For whatever it's worth, SoX automatically dithers the output only if it is less than 24 bits. I don't know what the reasoning was behind this decision.

 

[DonH56]

I have been pondering the question of dithering a 24 bit audio signal too.

Take the scenario of a convolution engine, processing a 16 bit, 44.1kHz input signal with some FIR filter(s).
If the engine works internally at 32-bit floating point by first normalizing its input signal to +/- 1.0, convolves the input and filter(s), then de-normalizes to a 24 (or 32 bit) integer for output (non-truncating rounding method), is there anything to be gained by applying dither to the output?

Me thinks not, but I would be grateful to have my notions reinforced or destroyed

Dithering is meant to reduce, or rather spread, quantization noise. In your example if everything were done at 32 bits right up to the DAC, the 16-bit DAC still has quantization noise at the 16-bit level, so yes you'd want to dither that.

 

[dc655321]

Dithering is meant to reduce, or rather spread, quantization noise. In your example if everything were done at 32 bits right up to the DAC, the 16-bit DAC still has quantization noise at the 16-bit level, so yes you'd want to dither that.

Thanks, Don.
Maybe I should have been more specific or I am not understanding you - I was assuming the DAC receiving the convolution-engine-abused 24 bit integers was capable of processing 24 (or 32) bit data. IOW, the engine output would NOT get truncated back to 16 bits for DAC consumption.

 

[DonH56]

Oops, you stated it was a 16-bit input signal which my little pea brain read as "16-bit DAC". Posting whilst working, bad idea. If there is a 24-bit DAC at the output then if the noise floor is sufficient perhaps you would not need to dither. However, dither is applied for several reasons and may benefit even when there is enough "other" noise around. Randomizing the quantization noise may help even at the 24-bit DAC level (while noting that no audio DAC I have seen really approaches true 24 ENOB performance, around 146 dB SNR with a noise floor around 216 dB), so maybe it does not matter.

 

[j_j]

Dithering is meant to reduce, or rather spread, quantization noise. In your example if everything were done at 32 bits right up to the DAC, the 16-bit DAC still has quantization noise at the 16-bit level, so yes you'd want to dither that.

No, actually, it slightly increases total quantization noise, but totally decorrelates it from the signal, making quantization a linear system with a noise source, rather than a nonlinear source.

That linearization is why it's so important on 16 (and less) bit signals.

And, yes, if you reduce to 16 bits at the output, indeed you )&(*& well dither!

 

[dc655321]

Posting whilst working, bad idea.

You are not joking. This whole working thing is really interfering with my ability peruse internet forums!

 

[DonH56]

No, actually, it slightly increases total quantization noise, but totally decorrelates it from the signal, making quantization a linear system with a noise source, rather than a nonlinear source.

That linearization is why it's so important on 16 (and less) bit signals.

And, yes, if you reduce to 16 bits at the output, indeed you )&(*& well dither!

May be semantics, or different definitions, since my data converter designs were not audio... I would (did) say (implied) dither increases the total noise, leaving out the word "quantization", because I treat quantization noise as a distinct type of noise. Dither increases total noise but quantization noise is a different thing to me, thus dither raises the noise floor, decreases SNR, and I agree decorrelates quantization noise from the signal (the desired result) while (because of) spreading the quantization noise spurs.

What I said earlier was "Dither (noise decorrelation) reduces (quasi-)correlated quantization noise at the cost of reduced overall SNR." Quasi-correlated because quantization noise is related to the signal but loose enough to be treated as random noise the vast majority of the time (not always for the systems I was working on, alas). Reduced SNR was due to the higher noise floor, but (again for the systems I worked designed which were not the same as yours) the added noise was white or colored noise. So in my world quantization noise itself did not increase though overall noise did. But that may be a function of the way we analyzed the noise and performed noise separation in our system. Or I may be all wet and not remembering the math. It involved Bessel functions and all that jazz so I've mercifully forgotten most of it... It is the same math that generates the actual spur content from quantization, and provides the 9N spurious-free dynamic range number.

And yes I know you know all this!

 

[j_j]

May be semantics, or different definitions, since my data converter designs were not audio... I would

..


And yes I know you know all this!

Um, I think we're agreeing past each other.

Strictly from a math point of view, what you are doing is linearizing the first and second order of the error in the quantization noise when you use TPD.

By that I mean that the resulting noise after quantization is not related to the input signal at all, AND that the energy (second order errror) is also not related to the input signal.

 

[j_j]

May I stir up the discussion a bit? NOS DACs need no filter because there always exist two filters: number one is the speakers/phones that shut off around 20 kHz. Number 2 are our ears.

Of course that won't prevent me from doing it right in the first place, but....you get it.

Your ears can respond nonlinearly, and our speakers, at least, don't cut off until 35000Hz. So no. Do it right.

 

[Angsty]

Is there a way to measure how well a DAC recreates recorded reverb? I got to thinking about this when reading an article by Herb Reichert in Stereophile about the R-2R Denafrips DACs' handling of reverb in recordings.

The one thing that appears to be consistent about subjective reviews of R-2R DACs is a sense of “soundstage width”, or “space”, or “three-dimensionally”. I interpret this as reverb. I’m not sure how the reproduction of this specific aspect in the recording is reflected (pun intended) in any one measurement.

Humans seem to have a strong attraction to and sensitivity to reverb. I wonder how this translates to the digital realm.

Any thoughts on this?

 

[SIY]

Is there a way to measure how well a DAC recreates recorded reverb? I got to thinking about this when reading an article by Herb Reichert in Stereophile about the R-2R Denafrips DACs' handling of reverb in recordings.

The one thing that appears to be consistent about subjective reviews of R-2R DACs is a sense of “soundstage width”, or “space”, or “three-dimensionally”. I interpret this as reverb. I’m not sure how the reproduction of this specific aspect in the recording is reflected (pun intended) in any one measurement.

Humans seem to have a strong attraction to and sensitivity to reverb. I wonder how this translates to the digital realm.

Any thoughts on this?

Yes, if it reproduces an input signal accurately, it will "recreate" recorded reverb. There's nothing special or magic about that part of the signal.