This page contains my views on how best to measure low level noise such as the self-noise of microphones, pre-amps and analogue-to-digital converts, or complete systems combining these.

This is an adjunct to the history page: ../#viewpoints .  Please see this page for the context of what follows.

I think the use of the CCIR 468 weighting curve is uncontroversial.  However, see below for reasons why, in the future, perhaps it would be better to use a curve which accentuates somewhat higher frequencies.

I think the use of the quasi-peak detector is clearly superior to using RMS, or worse-still, average metering, since noise is inherently uneven and human perception is particularly sensitive to higher level signals, even if they are very short.  Average metering gives very little precedence to narrow spikes.  RMS is better, but not as good as the quasi-peak technique.  It seems that the CCIR 486-2/3/4 approach to quasi-peak metering is a pretty good approximation to human perception - certainly better than RMS - when the noise contains peaks of energy which are narrow in time (short, impulsive).

It might be argued that some kinds of apparatus only generate noise which is so even as to make RMS just as good a measurement as quasi-peak.   My objection to this is that the suitability of RMS for a given measurement therefore depends on a subjective judgement about the nature of the noise being measured.

I think it would be better to use only the CCIR arrangements for measuring all low-level background and interferring noise in any system where the measurement is intended to reflect the subjective experience of humans listening to the resulting sound.

I don't accept the Dolby arguments about changing the reference level of the CCIR 486 curve to something other than 1 kHZ.  This was done simply to try to make the results more closely resemble the A-weighted RMS-metered results which are still the most common form of noise measurement. 

I think the CCIR 486 method has encountered some difficulties.  Hopefully, some of these will become historical:
  1. The only instruments which could perform the tests were expensive and quite rare.  There are no publicly available circuit diagrams, or DSP algorithms, AFAIK for the quasi-peak detector.

    Therefore, many companies and reviewers were unable to report the CCIR 468 noise figures.

    I intend to help resolve this by creating an open-source or public domain command-line program which can perform CCIR 468, A-weighted RMS-detected, and Dolby's so-called "CCIR-ARM" measurements on a .WAV file.
  2. "A-weighting" with RMS metering involves simple terminology which is, memorable and has meant the same thing since its inception decades ago. 

    The CCIR approach began as CCIR 468 (now known as the first edition CCIR 468-1) which did not have the crucial quasi-peak detector.  So just calling it "CCIR 468" is arguably too loose.  The editions -2, -3 and -4 are apparently the same for all practical purposes, so should we refer to it as "CCIR 496-4"? 

    That would be reasonable except for the fact that CCIR no longer exists and the formal title is now "ITU-R BS.468-4".  This technique has a marketing problem because it keeps changing its name.

  3. There may have been resistance in North America to proposals from Europe, especially from the work of researchers on the other side of the Iron Curtain.  This should be an historical curiosity now.

  4. The Dolby "CCIR-AMS" proposal made the adoption of full CCIR 468-2 much more difficult.  Ray Dolby's complaint about the quasi-peak metering circuit being expensive makes no sense at all to me.  It is not a complex circuit, and Dolby himself was a hot-shot designer.  BTW, no-one seems to have a patent on this quasi-peak metering system.
I think the CCIR 468-4 approach does a pretty good job of creating a single figure to predict the human perceptual impact of low-level random noise in an audio listening environment.  However, it is only valid as long as the noise is random and reasonably consistent over time.  There is no provision for handling situations where the noise worsens appreciably at times compared to other periods of time.

Nor does it cope with noise which has a significant tonal component.  I understand the human ear is particularly sensitive to higher frequency pure tones, such as 2 to 8 kHz.  A single tone, changing its volume, such as 3 times a second, is particularly noticable and annoying and, I think, would generate a CCIR 486-4 reading which significantly underestimates its propensity to annoy listeners.

With switching power supplies, the presence of cell-phones with high peak power outputs (GSM) and the integration of CPUs into many audio devices, narrow audio tones cannot necessarily be assumed not to exist. 

There is no longer a need for a filtering and metering system to be realisable in analogue circuitry.  It is safe to assume that all future test equipment is either implemented primarily in firmware, or consists purely of software to run on a PDA, PC or whatever.

This means it would be possible, in principle, to devise a far more sophisticated measurement system which still generates a single figure for "noise", but which could do a number of things which are beyond the scope of the CCIR 486 techniques:
  1. Use the latest psychoacoustical research (such as used for designing lossy audio compression algorithms) to more accurately estimate the impact of various signals, random and tonal, conistent and fluctuating.

  2. To automate the measuring process, such as by requiring a 10 second sound sample, and to produce figures which are directly related to the reference levels inherent in any digital audio file or signal.

  3. Potentially extend the technique into measuring the perceptual impact of distortion, as well as continual noise which is not correlated with the programme material.  For a given SPL playback level, as measured at the ear, this would make it possible for the measurement to estimate the impact of all noise and distortion products resulting from a given test signal, both as if the distortion and noise were being listened to alone, and in the presence of masking by the test signal.
All that is for the future, but it would be very useful when trying to estimate the noise and distortion inherent in microphones, pre-amps, ADCs, digitisation schemes, DACs and power amps.

For now, I think there should be continued effort to have the CCIR 486-4 approach more widely understood and used.

Freely available software will help.  However, I think there needs to be some centralised, or agreed, terminology and statement about this noise-measurment technique so manufacturers, reviewers and listeners can easily understand why it is superior to A-weighting with RMS detection, and why the figures it produces appear less favourable than those produced by A-weighting with any detector, or with any weighting filter with anything but the specified quasi-peak detector.

Such an explanation needs to be brief and contain informative graphics and proper references.  It probably needs to be a peer-reviewed paper which is widely supported by manufacturers and designers of test equipment / software.

This explanation probably needs to have its own website, with a simple URL which everyone can remember and refer to.  Then, noise figures can be referred to by using that URL alone.  So we need some kind of brief and distinctive name for this technique which will survive in perpetuity even if the standard itself finds itself in yet another standards organisation.