Laboratory Notebook 3 October 2002
Holes in White Noise
Today I am trying out an alternative to the CoolEdit FFT NR process.
This is so good at picking voices out of the mud that for a while I have wondered if in fact it was doing an EVPMaker operation on the raw input. But no -
The problem is getting NR done on-line - simply by filtering or whatever.
To find out what the CoolEdit NR is in fact doing I developed the following strategy.
I took a sample of raw audio from an Alpha being operated, and I chose a portion where the artefacts are prominent. I used this noisy selection to produce the Profile for Noise Reduction.
Ideally, one would be able to look at this Statistical Profile and design a bank of filters accordingly - but it is shown in a little window that does not show very much.
So I saved that Profile. Next I took some seconds of pure white noise. When you look at the frequency analysis after a scan then you will see a flat horizontal line. In other words the white noise has a flat spectrum. As it should be.
I now wanted to see what holes the NR using the saved Profile punched in that flat-line.
So I did a noise reduction on the white noise sample using the Profile previously saved, and looked at the spectral response after a scan. The "holes" that now showed in that spectrum revealed where the CoolEdit NR filters had been placed.
I then took that spectral response and mimicked it as a complex FFT Filter.
Now, theoretically, if I was to pass the raw audio through the newly designed filter it should yield the same result as doing a Noise Reduction using CoolEdit with the saved Profile.
If you did that and then looked at the spectral responses of each you would see that they were close - but listening - though improved the simply filtered version was still not as good as the CoolEdited one.
Further tests showed that what the Cool people dont say - or if they do say it then I haven't found it - what they dont say is the their NR system also includes what seems to be Noise Gating and Companding.
I have now put in some of that - resulting in a further improvement - but there is still a gap.
There seems to be a nice vocoder-style hollow resonance to the NR results that is not present from the more erratic simply filtered result. Possibly a final lumpy broad-band mid-frequency filter may help. I am reflecting on the fact that with appropriate acoustic design all this was taken care of in the past much more simply. To get away from any points of criticism I missed out the acoustic section of the older Mk 3.5 and earlier models, concentrating instead on a purely electronic model