Various noise reduction systems have been designed on the multi-band Downwards Expander principle. When a single-band Downwards Expander, as described last month, encounters a very quiet signal, it simply reduces the volume; no matter whereabouts in the audio spectrum the wanted signal is, the audio level of the whole signal is turned down. What might work better is a system that can treat separate sections of the audio spectrum independently -- ie. a multi-band system.
Such a system may be implemented by separating the audio spectrum into a number of bands, and then treating each of these bands with its own downwards expander. Consequently, the device can be reducing the volume in one frequency band while leaving the signal in other bands untouched. This can lead to impressive results; devices such as the 5-band Roland SN550 put this theory into practice.
As you might imagine, the more bands you can process, the more effective the system is likely to be, but it is technically difficult to construct the appropriate filters using analogue circuitry. Digital technology, where the filters can be made extremely accurate and very selective, is far more adept at this kind of job. Consequently, a multi-band unit of this type will cost more than the simple Dynamic Noise Filters described last month.
Roland's SN550 also incorporates a hum filter to remove low frequency noise and its associated harmonics, which might spread far into the upper reaches of the audio spectrum. Again, this is made possible by the use of powerful digital filters, which create very narrow notches so as not to compromise the wanted signal. Roland's filter can also be set to track the incoming mains frequency, ensuring that it is always precisely tuned. A manual tuning control is also provided, which is useful for removing hum from taped recordings where there may have been tape speed changes.
This type of noise reduction has now found its way onto computers. Modern digital audio workstations (usually hard disk editors with other functions added on) utilise processor chips known as Digital Signal Processors (DSPs), which can perform millions of calculations every second upon digitised audio data. These systems also split the audio spectrum into multiple bands and apply downwards expansion in the digital domain.
But even multi-band units have no way to make a true distinction between wanted signal and unwanted noise; they still act upon the general assumption that, if the signal level approaches its noise floor, all that is present is broadband noise. Consequently, even the most sophisticated (and expensive) Downwards Expanders and Dynamic Filters inevitably remove some of the genuine signal. The consequences of this are well understood and, to a greater or lesser extent, unavoidable -- loss of high frequencies, loss of ambience, and degradation of hard transients.
Finally, we arrive at the most sophisticated noise removal technology yet implemented in available systems: Spectral Subtraction. But first, some simple mathematics...
All the previously discussed noise abatement techniques use filters, gain controls, or a combination of both to achieve their results. Whether implemented in the analogue or digital domains, all such filters and gain controls are 'ratio' devices -- that is, if, at any given frequency, you remove half the power of the noise, you also remove half the power of the signal at that frequency. Of course, this only applies to signals falling below the selected threshold, but because sub-threshold signal and noise is treated in exactly the same manner, audible side-effects can show up.
What more can be done using digital processing? Imagine a signal that has, at a given frequency, 100 units of 'volume' on some arbitrary scale. Let's also say that, by measuring the noise content of the signal during an otherwise silent moment, you have determined that there are 20 units of noise present on the same scale. It should be possible to remove the noise contribution by subtracting these 20 units in the digital domain. If, a moment later, the 'volume' of the signal drops to 40 units the digital process is still able to remove the full 20 units of noise. No analogue device can precisely emulate this 'subtractive' filter, and herein lies the power of the computerised noise reduction system. Given enough bands, this type of process can tackle excessive noise problems with a high degree of success.
Unfortunately, such power does not come cheap. If you have a few thousand pounds to spare you can acquire a Digidesign Pro Tools system with DINR (Digidesign Intelligent Noise Reduction) software, or for a cool £30,000 you can pay off your mortgage, buy your other half a Porsche, or invest in a British CEDAR System (developed at Cambridge University) or an American NoNoise system. CEDAR and NoNoise are IBM and Macintosh (respectively) computer-based systems that, among other things, can split the audio signal into 1024 bands (each a mere 22Hz wide) and apply Spectral Subtraction to each of these. Splitting the signal into so many bands means that you can be very precise about how much noise you remove, subtracting a lot of noise at (say) 8kHz, while leaving 8.1kHz virtually untouched. If this sounds too good to be true, in some ways it is.
The noise spectrum of a recording (the sonic 'fingerprint') can only be accurately measured if there is an otherwise silent passage within the music. If the fingerprint is wrong (maybe because you have captured some lingering reverb, or because the original recording engineer has faded sections in and out of the recording) the amount subtracted will be wrong, leading to some very unpleasant sounding side-effects. And, just to make matters worse, many tracks are 'close edited' -- the run-in and run-out of the track have been removed -- making it impossible to take a fingerprint.
Assuming that you have managed to obtain a perfect noise fingerprint, you might expect to produce a very good restoration of your track with little or no side-effects. Yet experience shows that all attempts to use an unmodified noise fingerprint lead to a dry and dull sounding result. This is because the fingerprint is merely a snapshot of the noise content of the material, accurate only at the instant at which it is taken. The very essence of noise is its random nature, and because the profile of the noise content is constantly changing, it is necessary for the noise fingerprint used by the system to change also.
A new approach to this problem has been incorporated into CEDAR Audio's HISS-2 system. With a noise fingerprint that is updated 44 times per second (allowing CEDAR to track variations in the noise content of the recording); new algorithms which 'look ahead' at the incoming signal, responding to transients before they occur; and an ambience control which ensures that sounds are not prematurely cut short, CEDAR combines many of the analogue and digital ideas discussed above. This means that, in theory at least, the amount of noise being removed is always appropriate. The outcome is that HISS-2 enables the right amount of noise to be removed without damaging the source signal.
Because of the very high cost of such systems, you'd be unlikely to buy one on the off-chance that you need to use it. More likely you would hire time on a system, with an experienced operator, when required.
While you should never think of noise removal as a retrospective justification for sloppy recording, there will always be some residual noise inherent in the recording equipment, the sound processors, the instruments being recorded, and so on. For as little as £30 spent on a second-hand noise gate footpedal, you can begin to clean up your act. For about 10 times that much, you could invest in a Dynamic Noise Filter which, used sparingly, can make a significant contribution to your aural hygiene.
Another factor of 10 buys you a more powerful multi-band device, such as those used by the radio and TV industries; and a final factor of 10 gets you into CD and film soundtrack mastering quality. In an ideal world, it would be nice to have access to the power of a system such as CEDAR or NoNoise for the price of a noise gate. Currently, this isn't possible, but technology has a habit of percolating down the price scale...
The multi-band expander is equally useful for noise suppression in mastering, and for live broadcasting. Perhaps a little expensive for the home studio, the multi-band may well be worth hiring for mastering sessions -- especially if your master tapes are rather more hissy than you can live with. They are also useful for cleaning up old analogue recordings while archiving to DAT. Models with digital hum filtering can work miracles with buzzy guitar tracks. Multi-band computer-based systems are generally capable of better results than stand-alone units having only a few bands. Unless used sparingly, these units can cause audible side-effects.
At the more affordable end of the scale, Digidesign's DINR software uses a form of Spectral Subtraction based on a fixed noise fingerprint and can produce superb results, so long as the level of noise reduction is limited to around 6dB. If more noise removal is attempted, audible side-effects can become evident.
The more advanced Spectral Subtraction systems (CEDAR and NoNoise being the market leaders) are the preferred choice for top mastering studios and post-production houses worldwide. They can be used to clean up new masters that exhibit noise problems, and may also be used to restore old recordings for re-release. Such systems are extensively used in film soundtrack work, where any residual noise is unacceptably obtrusive.
AMPLITUDE: The amplitude of a signal defines its loudness (or volume). A loud sound has a high amplitude, while a quiet sound has a low amplitude. Amplitude can also be defined at a given frequency (see below). The most common unit of amplitude is the decibel (dB).
AMPLITUDE RESPONSE: Certain audio devices alter the amplitude of a signal according to a defined response curve (see diagrams). Thus the amplitude of the output signal will be different to that of the signal before it passes through the device.
ATTACK: The speed at which an audio processor reacts. Fast attack settings enable the device to change response rapidly. Slow attack settings force the device to take a perceptible time before the final response is obtained.
AUDIO SPECTRUM: Measured in Hertz (Hz), or 'cycles per second'. The limits of human hearing are often quoted as being 20Hz to 20kHz, although these extremes will differ according to the age of the listener and the volume of the sound. A high frequency sound (such as a whistle or a hi-hat) has a high pitch; while a low frequency sound (bass drum, double bass, etc) has a low pitch. High and low frequencies should not be confused with high and low amplitudes.
FILTER: Strictly speaking, any device which alters an audio signal is a filter. More commonly, however, the name 'filter' is limited to those devices which change the frequency response of a signal: a low-pass filter removes high frequencies from a signal; a high-pass filter removes low frequencies from a signal. The starting point for these filters is called the 'shelf frequency'. Other types of filter include band-pass, band-reject, and notch filters.
HYSTERESIS: Most audio devices offer an essentially 'linear' response, that is, if a certain condition is reached, the device always offers the same response. Some units, however, have hysteresis responses: if the condition is approached from one side (eg. the original signal is decreasing in volume) the action of the unit is different than if the original signal is increasing in volume).
RELEASE: The speed at which the audio processor returns to its 'idle' setting after the requirement to produce a response has disappeared.
ROLL-OFF: The response of a low-pass or high-pass filter is described in 'decibels per octave' (dB/oct). A low-pass roll-off of 6dB/octave states that at twice the shelving frequency the amplitude of the output signal is half that of the input signal. Essentially, the more dBs per octave, the sharper the filter response.
SIGNAL TO NOISE RATIO: This is the difference between the amplitude of a sound and the amplitude of any unwanted accompanying noise. If a signal of total amplitude 100dB contains 40dB of broadband noise, we can say that the signal to noise ratio (S/N) is 60dB.
TRANSIENT: Many sounds may be thought of as combining an energetic initial burst with a longer sustained portion. The initial burst, which often has a higher amplitude than the rest of the signal, usually contains most of the high frequencies in the sound, and is generally of short duration. Hence the name 'transient'.