Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[Barry Hufker]

I've never really thought about this topic until recently.

When manufacturers of A/D (or even D/A systems I guess) converters specify noise, do they mean Peak or RMS?  And if so, what is the testing method? And what are reasonable figures for a typical converter at 24 bits and 96k?

I've begun to wonder about the nature of all A/D converters...

[John Roberts {JR}]

We can wait for a more comprehensive answer but in general modern convertor noise is measured as RMS or average. A little quick math will reveal that it is higher than the quantization floor as predicted by the digital word length.  This noise will self dither the lower bits and give the conversion a more analog sounding noise floor.

JR

[Barry Hufker]

Thanks JR.  What surprises me not so much is that the noise is measured RMS, but no one talks about peak noise, which is about 20 dB higher.  I'm not sure of the implications of peak noise but it is amazing that readings such as -72 dB for 16 bit or -82 dB for 24 bit occur.  If these figures are correct (and at the moment I believe them to be) and "universal" then this is a revelation to me.

[John Roberts {JR}]

Barry Hufker wrote on Sun, 29 June 2008 10:21

Thanks JR.  What surprises me not so much is that the noise is measured RMS, but no one talks about peak noise, which is about 20 dB higher.  I'm not sure of the implications of peak noise but it is amazing that readings such as -72 dB for 16 bit or -82 dB for 24 bit occur.  If these figures are correct (and at the moment I believe them to be) and "universal" then this is a revelation to me.

RMS generally correlates better with perceived loudness than peak. Peak is more important in the context of saturating full scale than when down several tens of dB (IMO).

JR

[Barry Hufker]

JR, I fully accept your thought about RMS, still I'm surprised that peak noise levels are that high.  While you're saying RMS is closer to what we hear, doesn't the peak level seem to undermine the notion of a "noiseless" digital process?  People put forth numbers such as -107 or -114 (A weighted), etc.  yet the noise can be (and I guess in a certain sense is) much higher.

Am I still missing the point?

[johnR]

True white noise is essentially random, meaning it can be any value within the amplitude and frequency limits of the system. That's a simplified explanation and ignores the fact that some values are more likely to occur than others, but the important thing is the noise voltage at any particular instant is unpredictable. The only way to get a meaningful measurement is to take the RMS value.

Measuring noise with a peak meter is likely to give a pretty strange result because there will be peaks that are much higher than average occurring at random times. Depending on the ballistics of the meter, these will be averaged out to some extent but the result won't necessarily tell you much.

[Barry Hufker]

Any thought then as to how these peaks affect signal quality?  How do they affect quantization?  I guess I'm wondering... there's got to be some qualitative effect right?  Even if the peaks themselves even out?  They are 20 dB higher than RMS...

EDIT: I hope I'm not being redundant or dense...

[johnR]

Barry Hufker wrote on Sun, 29 June 2008 20:29

Any thought then as to how these peaks affect signal quality?  How do they affect quantization?  I guess I'm wondering... there's got to be some qualitative effect right?  Even if the peaks themselves even out?  They are 20 dB higher than RMS...

Whatever the effect, there's no escaping it. The sound reaching the microphone is noisy due to random motion of the air molecules in the signal path. The random motion of air molecules against the microphone diaphragm produces more noise. In the electrical domain, everything with a resistance produces noise, the higher the resistance the higher the noise (this is why it's important to terminate inputs with the correct source impedance when measuring noise in audio equipment).

[Barry Hufker]

I understand what you're saying -- noise is everywhere and there's no escaping it.  But I would also take that to mean if we can reduce the noise inherent in a system then the spikes in A/D converters will be lower (for instance if we can ever get true 20 to 24 dB conversion.  The noise will be there only much lower than at present.

I suppose one last statement would be what I said earlier, I'm surprised digital systems aren't currently less noisy than I was naive enough to believe.

[johnR]

I'm not an expert on sources of noise in A/D converters, but this article looks quite informative:

http://www.analog.com/library/analogdialogue/archives/40-02/ adc_noise.html

[John Roberts {JR}]

Barry Hufker wrote on Sun, 29 June 2008 15:21

I understand what you're saying -- noise is everywhere and there's no escaping it.  But I would also take that to mean if we can reduce the noise inherent in a system then the spikes in A/D converters will be lower (for instance if we can ever get true 20 to 24 dB conversion.  The noise will be there only much lower than at present. I suppose one last statement would be what I said earlier, I'm surprised digital systems aren't currently less noisy than I was naive enough to believe.

I am not the converter expert here, but that hasn't shut me up so far.

I'm not sure I would characterize the peak value of the noise as spikes. The random noise, sometimes lines up and most times doesn't.

FWIW there is usable signal reproduction below this apparent noise floor, just like you can hear signal down in the residual noise of magnetic tape, you can hear signal below this converter noise.  

More is always better, right? I'm sure there are engineers working feverishly to be the best on paper, to better land the "price is no object" customers. OTOH I saw 24b codecs offered by a Chinese fab for $0.90 in quantity (Cool Audio).

I'm not sure it matters much in the context of typical audio signals we are called on to process or reproduce. As long as the channel poo is 10+ dB below the signal poo... and random, no worries.

JR

[Barry Hufker]

Thanks John,

You're a converter expert as far as I'm concerned.  Your answers have been direct and very helpful.

And I've got a new phrase - "signal poo".  I may change the name of my company to that.  Who could resist recording at Signal Poo?

Signal Poo™ John Roberts, 2008.

I'm going now to have a look at the article.

Thanks!

[bruno putzeys]

Yup, you can't characterise truly random gaussian noise by the peak level because it's a statistical thing. The maximum peak to be expected, with a given odds, in any given time interval, increases with the logarithm of the length of that interval. This can lead to an amusing visual artefact when plotting random noise on a logarithmic X scale, as the thickness of the line covers an increasing length of time.

You'd swear the noise is growing with time but in reality, as each pixel represents more and more data points, the odds of hitting a greater maximum increases.

Anyhow, this is not different from noise in analogue circuits. The only reason why it's more obvious with converters is because digital peak meters have a "zero" attack time while analogue PPM's still have a few ms of response time.

The matter is also touched on in the Lavry Gold manual.