Understanding Noise Weighting Curves

Microphones, amplifiers, and recording systems all add some residual noise to the signals passing through them, but the noise generating mechanism and so the spectral content of the noise is different in each case. Tape recorders, for example, add a lot of high frequency noise, compared to the more even spectrum of some sources, and this is particularly audible because the human ear is very sensitive in the 5 -7kHz region.

If noise measurements are to have any real value, allowing fair comparisons of the true noise contributions from different types of equipment, they should give a figure that is representative of what we hear, and the first step towards this is the use of a Weighting Filter , which emphasises some frequencies more than others. Additionally, it turns out that there is another property of noise that varies, which also affects its perceived loudness, and that is its amplitude distribution or 'peakiness'. Some noise spends much of its time around a mean level, without significant peaks, but other sources of noise contain short peaks which the ear seems to take special notice of, the loudness being more related to the peak value than the mean. Just as very high frequencies are barely heard, so very brief peaks also go unheard, and so just using a peak reading rectifier is not the answer. Instead a 'Quasi Peak' or 'almost' peak reading meter is needed, its 'attack' time tailored to the ears behaviour.

Early attempts to improve on basic noise measurement used what came to be referred to as the 'A' Weighting curve, which had been derived from the work of Fletcher and Munsen (1933) who investigated hearing variation with frequency and found that they needed to plot a set of 'equal loudness' contours because the ears frequency response was different at different loudness levels. The curve representing loudness equal to that at 1kHz and 40dB SPL (sound pressure level) was known as the 40-phon curve, and was (supposedly) adopted for weighting purposes. Later work by Robinson and Dadson (1956), refined the method producing significantly different, and more accurate curves however. In 1968, two inventions began to make clear the inadequacy of A-weighting; the introduction of FM Radio and the Compact Cassette. The latter in particular, was found to sound a lot less noisy (10dB) with Dolby noise reduction switched on, without measuring significantly better, and this led to work on better ways of measuring noise. The BBC (British Broadcasting Corporation) undertook a research project, culminating in BBC Research Dept Report EL-17 entitled "The Assessment of Noise in Audio Frequency Circuits", in which they studied the effectiveness of various wieghting curves and rectifiers that had been devised by the world's broadcasters, on all sorts of extreme noise sources, and they chose one combination as being very effective. Later work refined this into what became a world standard known as CCIR468, which, though it has gone through versions 1 to 4 is still essentially unchanged, only the permitted tolerances being altered. Unfortunately, though CCIR468 was adopted by many major broadcasting organisations, and also incorporated into standards from BS (British Standards) and IEC (International Electrotechnical Commission) to EBU (European Broadcasting Union), and for a while even became championed in consumer reviews of Cassette Decks (notably by Angus McKenzie in the 1970's) it is now falling into disuse for one simple reason: in an age when advertising rules, big figures are reckoned to sell products, even if they are meaningless, and A-weighting gives bigger numbers! This is something I am very keen to put right, before audio quality measurement falls completely into disrepute. Measurements ARE valid, but only if they are properly weighted, and CCIR weighting works very well indeed. Lets all use it!

Noise Weighting Curves Compared

The green curve is taken from the equal loudness contours of Robinson and Dadson (1956) and represents the 40 Phon level.
Note that although the A-Weighting curve (Red) is often supposed to approximate to this, it actually differs very significantly, being 10dB low at 100Hz and 6db low at the peak of the all-important 2 to 6kHz region where noise in audio systems predominates perceptually. It also fails badly in the cutoff region being roughly 25dB high at 20kHz! Although it has been pointed out that the A-weighting approximates better to the earlier Fletcher-Munsen contour, on which it was based, this does not alter the fact that it is a poor approximation to the later and more reliable curves. As well as reflecting the innaccuracies of very early headphones and microphones, the Fletcher-Munson curves suffered another deficiency, becuse they did not use frontal sound source. We now know that the external ear introduces big variations in frequency response depending on the direction of the sound source, and its purpose is to introduce such variations as directional cues that the brain can interpret. Sound through headphones should never be presented with a flat response, but should attempt to correct for this effect, as most modern headphones attempt to do (though badly).
The ITU-468 weighting curve (Blue) (originally CCIR 464) arguably comes closer in this region, the exact nature of the notch at 8kHz being so listener dependant as to be not worth rendering. I have argued that this curve might theoretically be expected to be higher at high frequencies and lower at low frequencies because it is based on the ears sensitivity to noise; and the ears hair cell channels, being of wider absolute bandwidth at higher frequencies, give higher sensitivity to high frequency noise. If the ear had geometric channel spacing (which it does not) then a 3dB per octave tilt would have to be added to the R-D curve for noise.