The term 'crackle removal' is slightly misleading because the CEDAR crackle removal algorithms remove a range of high density, small amplitude, short duration, additive impulsive disturbances. These can cause a variety of degradations with differently percieved sounds.
If these disturbances are randomly distributed, the degradation will sound like crackle or a 'chip-fryer' type of noise. For example, there is a fungus that eats the vegetable matter contained in old 78rpm records. It leaves millions of pock-marks on the surface of the disc, and these create impulsive disturbances in the replayed waveform, thus giving 78s their characteristic crackly surface noise. If the impulsive disturbances are regularly spaced, the degradation will sound like a buzz. For example, the controllers used to vary the light intensity in lighting racks do so by cutting out part of the AC power twice every cycle. This process creates sharp electrical transients. If nearby audio equipment is poorly shielded these transients may induce regular clicks that sound like buzz. If the impulsive disturbances are correlated with the signal, the degradation will sound like distortion. For example, mild clipping distortion can be considered as an impulsive disturbance added to a pre-clipped original signal, such that the result stays within the available headroom.
It is important to note that there are many other types of distortion (such as tape saturation) that the crackle removal algorithm can not address. This is because the natures of these other distortions are not caused by the high density impulsive disturbances described above.
Crackle and distortion exhibit a broad spectral envelope, so people have tried to use spectral dehissing techniques to remove them. In contrast, buzzes tend to have regular spectral peaks, so the preferred technique (up until now) has been to apply a large number of notch filters, each of which removes one peak of the envelope plus any signal that also happens to exist at that frequency. But each impulse is highly localised in time and, at most, 10% of the signal may be affected by the impulses. Consequently, the other 90% of the signal is undamaged,so filtering, or any other process that affects the undamaged signal, is wholly inappropriate.
Unfortunately, interpolating each impulse using a declicker is also not satisfactory because of the fundamental nature of the crackle. A standard click-detection algorithm will often miss the impulses, and it will fail to cope with the high density of those impulses. Furthermore, declick algorithms assume that a click or scratch has totaly corrupted the audio and that there is, therefore, no useful information about the signal during the click. The disruptions that cause crackle are much smaller and additive , so this assumption is no longer appropriate. Declicking a crackly signal discards a lot of useful information that an optimal decrackle algorithm can use to improve the sound quality of the processed audio.
CEDAR's research into developing an optimal decrackle process has shown that a crackly signal may be divided into two components. The first of these contains the bulk of the clean (desired) signal. The second (the 'split' signal) contains all the degradation plus the residue of the clean signal.
Once the split signal has been obtained, the crackle detector has a much better chance of identifying the undesired disturbances. Following detection, a suitable interpolator optimised for the high-density nature of the crackle can then be used to remove the disturbances from the split signal. This does not affect the body of the audio signal. Once the split signal has been interpolated, it may be recombined with the body signal in order to recreate the original signal, but without the crackle. By doing this, we have retained the useful information contained in the original signal, even during the period of the interpolation.