Introduction to noise measurement
Noise in general refers to unwanted sound, often loud, but in audio systems it is the low-level hiss or buzz that intrudes on quiet passages that is of most interest. All recordings will contain some background noise that was picked up by microphones, such as the rumble of air conditioning, or the shuffling of an audience, but in addition to this every piece of equipment which the recorded signal subsequently passes through will add a certain amount of electronic noise, which ideally should be so low as to contribute insignificantly to what is heard.
Microphones, amplifers and recording systems all add some electronic noise to the signals passing through them, generally described as hum, buzz or hiss. All buildings have low-level magnetic and electrostatic fields in and around them emanating from mains supply wiring, and these can induce hum, commonly at 50 or 100Hz into signal paths. Screened cables help to prevent this, and on professional equipment, where longer interconnections are common, balanced signal connections (most often with XLR connectors) are usually employed. Hiss is the result of random signals, often arising from the random motion of electrons in transistors and other electronic components, or the random distribution of oxide particles on analog magnetic tape. It is predominantly heard at high frequencies, sounding like steam or compressed air.
Attempts to measure noise in audio equipment as rms voltage, using a simple level meter or voltmeter, do not produce useful results; a special noise-measuring instrument is required. This is because noise contains energy spread over a wide range of frequencies and levels, and different sources of noise have different spectral content. For measurements to to allow fair comparison of different systems they must be made using a measuring instrument that responds in a way that corresponds to how we hear sounds. From this, three requirements follow. Firstly, it is important that frequencies above or below those that can be heard by even the best ears are filtered out and ignored; by bandwidth limiting (usually 22Hz to 22kHz). Secondly, the measuring instrument should give varying emphasis to different frequency components of the noise, in the same way that our ears do; a process referred to as weighting. Thirdly, the rectifier, or detector, which is used to convert the varying alternating noise signal into a steady positive representation of level should take time to respond fully to brief peaks to the same extent that our ears do; it should have the correct dynamics.
The proper measurement of noise therefore requires the use of a specified method, with defined measurement bandwidth and weighting curve, and rectifier dynamics, and two main methods defined by standards are currently in common use: A-weighing, and ITU-R 468, formerly known as CCIR weighting.
A-weighting uses a weighting curve based on equal loudness contours that describe our hearing sensitivity to pure tones, but it turns out that the assumption that such contours would be valid for noise components was wrong. While the A-weighting curve peaks by about 2dB around 2kHz, it turns out that our sensitivity to noise peaks by some 12dB at 6kHz. Another weakness of A-weighting is that it is usually combined with an rms (root mean square) rectifier, which measures mean power, with no attempt made to account for proper hearing dynamics.
ITU-R 468 weighting
When measurements started to be used in reviews of consumer equipment in the late 1960s it became apparent that they did not always correlate with what was heard. In particular, the introduction of Dolby B noise-reduction on cassette recorders was found to make them sound a full 10dB less noisy, yet they did not measure 10d better. Various new methods were then devised, including one which used a harsher weighting filter and a quasi-peak rectifier, defined as part of the German DIN45 500 Hi Fi standard. This standard, no longer in use, attempted to lay down minimum performance requirements in all areas for High Fidelity reproduction.
The introduction of FM radio, which also generates predominantly high-frequency hiss, also showed up the unsatisfactory nature of A-weighting, and the BBC Research Department undertook a research project to determine which of several weighting filter and rectifier characteristics gave results that were most in line with the judgment of panel of listeners, using a wide variety of different types of noise. BBC Research Department Report EL-17 formed the basis of what became known as CCIR recommendation 468, which specified both a new weighting curve and a quasi-peak rectifier. This became the standard of choice for broadcasters worldwide, and it was also adopted by Dolby, for measurements on its noise-reduction systems which were rapidly becoming the standard in cinema sound, as well as in recording studios and the home.
Though they represent what we truly hear, CCIR weighted noise figures are typically some 11dB worse than A-weighted, a fact that brought resistance from marketing departments reluctant to put worse specifications on their equipment than the public had been used to. Dolby tried to get round this by introducing a version of their own called CCIR-Dolby which incorporated a 6dB shift into the result (and a cheaper average reading rectifier), but this only confused matters, and was very much disapproved of by the CCIR.
With the demise of the CCIR, the 468 standard is now maintained as ITU-R 468, by the International Telecommunications Union, and forms part of many national and international standards, in particular by the IEC (International Electrotechnical commission), and the BSI (British Standards Institute). It is by far the best way to measure noise, and the only way that allows fair comparisons; and yet the flawed A-weighting has made a comeback in the consumer field recently, for the simple reason that it gives the lower figures that are considered more impressive by marketing departments.
Signal to noise ratio and Dynamic range
Hi-fi equipment specifications tend to include the terms signal to noise ratio and dynamic range, both of which are confusing and best avoided. Noise has to be measured with reference to something, but this should be alignment level. Signal to noise ratio has no real meaning as audio signals are constantly changing so there is no such thing as signal level. Dynamic range used to mean the difference between maximum level and noise level, but maximum level is often hard to define, for example on analog tape recordings, and the term has become corrupted by a tendency to refer to the dynamic range of CD players as meaning the noise level on a blank recording with no dither, in other words just the analog noise content at the output. This is not particularly useful; especially since many CD players incorporate automatic muting in the absence of signal to make them appear even quieter!
Professionals measure noise in dB below alignment level, which is a reference point above which headroom exists up to maximum permitted level. Professionals often allow 18dB of headroom, as recommended by the EBU (European Broadcasting Union), so a noise level of 60dB ITU-R 468 would represent a dynamic range of 78dB, which if measured A-weighted might come out 11dB better at 89dB. A noise level of -60dB AL would be considered reasonably good by professionals, with 68dB representing the best attainable from 16-bit digital audio (noise shaped), and more than good enough for most purposes.
Audiophiles may talk in terms of 96 to 120dB dynamic range, but they often fail to refer to any measurement standard, making the figures meaningless. Attempts to calculate the dynamic range of digital audio on the basis that 16 bits represents a ratio of 65000:1 or 96dB are invalidated by the fact that the full digital count represents the peak possible level, rather than the rms equivalent of the maximum possible sinewave, while the minimum count of one has little to do with the noise level, which depends on the type of dither (or noise-shaping) used. They also fail to take any account of weighting for subjective validity.