CEDAR Single-Ended Noise Reduction System (part 5)
The second level of operation is the 2-pass process which can be done after Descratching. This will remove all of the remaining impulsive noises. How it actually works, so I understand, is too mathematical to be fully comprehended by anyone with less than an infinite IQ, but Figure 1 gives a reasonable idea of the process. First the signal is split into two parts, one which contains 90% of the signal and no noise, the other has the remaining 10% of the signal and all of the noise. The Splitter level, as it is called, is set by the operator while listening to the signal in real time. It is not a critical operation as long as all the noise goes into the portion to be processed. In this portion, the noise is very much more prevalent than the signal so a descratch process, similar to the original Descratch, can be applied to root out nearly all the noise. This final descratching is incorporated into the Recombine procedure, which puts the two parts of the signal back together, hopefully resulting in all the signal that there ever was but with none of the noise. Like Descratch, there are two adjustable parameters, Order and Threshold.
After the 1-pass and 2-pass scratch removal procedures there should remain only the signal and non-impulsive noise. Its probably fair to say that up to this point CEDAR is purely attempting to restore the signal to its original state.
The next step along the road to enhancement is Dehiss, where the operators judgement is critical to the results from the process. Fortunately, Dehiss works in real time, so given a keen ear and a good acoustic environment it should be possible to find an optimum point where all the bath water has been thrown away but the baby has been wholly left behind.
Firstly a fingerprint is taken of the noise spectrum, from a section of the recording which is without signal. Then, during processing, the software decides on a moment to moment basis whether or not noise is present, and if it is, it is removed. Note that this is not done by equalising with the inverse of the noise curve. EQ in CEDAR is a separate process. There are two main parameters which the operator must adjust correctly: the Multiplier which sets the decision level at which CEDAR will say This is probably noise so Ill get rid of it and the Attenuation which defines the residual noise level after processing. Roughly speaking, the Multiplier adjusts the effect between hissy and dry, and the Attenuation adjusts between warbly and sounds OK. The range of defects it is possible for the process to produce is some way removed from the conventional boundaries of audio malaise so description is difficult, but for any recording there should hopefully be a combination of parameters that has a solely beneficial effect.
Once Dehissed, the recording can be Equalised with CEDARs 512 band EQ (Im afraid I didnt count them). Referring back to Figure 1, an interesting feature is the ability to split the signal into 90% signal/no noise and 10% signal/100% noise components. If the portion that contains only the signal is equalised, then it is apparently possible to get 30dB to 50dB of noise free equalisation!