CE 2.0 & 2.1, Audition 1.0 & 1.5 FAQ Comments

Post your comments here.

Rather than use up the forum resources, you might consider sending me an e-mail and I will be glad to respond.  I have not posted many comments on this - or previous - forums, but I would be more that willing to share my 10-plus years restoring old vinyl LPs and magnetic tapes.

The purpose of forums of this nature is to spread information, and make it freely available - that's exactly what the resource is intended for. It runs somewhat contrary to the spirit of the forum, therefore, to only offer to do this privately, I'd say.

In terms of restoration, I think we've actually covered a vast  amount of ground over the last decade - it's worth checking out the archive thoroughly. And we have a fair few professional restorers as forum members too...

I have yet to "master" the art of using the noise reduction feature of CoolEdit2k.  I understand the principles of the noise reduction algorithm and have learned that even the "best" noise sample - sans desirable audio - can be applied either too much or too little.  I know there are "old hands" that acquired the "knack" of noise reduction and I would appreciate some advice.  The settings I have settled on are: NR. of snapshots in profile (2048).  Noise reduction level (I use about 52).  FFT size (I use 2096).  Reduce by "X" db; this has me stumped.  I usually apply about 40 dB but sometimes that seems too much.  I use the default precision factor of 8, smoothing amount of 1 and transition width of zero dB.  Can someone enlighten me as to how these settings control the amount of the noise sample that is subtracted from the waveform of interest?  And, if I have gone "wrong," how should I "fix" it?  (Alas, ten years of using CE2k has not taught me very much...) - R Neff.

Reduce by "X" db; this has me stumped.  I usually apply about 40 dB but sometimes that seems too much. 

Try reducing this to 10 or 12dB. If you need more then do a second pass. I also tend to use an FFT size of 8192 but larger FFT's can take up more processing power.

Cheers

James.

Thanks James from NH, I will increase the FFT size to 8192.  The size/time to execute is not a problem on this computer.  The parameters about which I am uncertain (i.e., clueless) is the amount of noise reduction level to apply.  I generally use about 50 or so, that coupled with the "reduce by" in dB.  The default is 40, which I have learned can be excessive and is, of course, coupled with the noise reduction level.  I guess one must experiment with these two parameters to find out which is best for a given application.  Thanks again for the response, although I had to "dig" to find your post.  Someday I may get the hang of this new fangled forum...

Thanks James from NH, I will increase the FFT size to 8192.  The size/time to execute is not a problem on this computer. 

In that case set it to the absolute maximum value, and use your ears. Everything becomes less critical in terms of settings, and unless you have an awful lot of very low frequency noise, the results will improve dramatically.

OK, Steve!  Well, let me see.  When all else fails, try following instructions...  Back in the "good ol' days" when Steve Fazio was lurking around the old Synt forum, he once said, "Lower FFT sizes of 2048 and below have better response times (I interpreted that as computer execution time) but poor frequency resolution (which makes sense).  Higher FFT sizes of 12000 or more (!) have very good frequency characterization of the analyzed signal but may introduce "whooshing" and reverb-like sounds."  I did use James suggestion of an FFT size of 8192 and noticed a marked improvement of the characterization of a noise sample.  I arbitrarily set the number of samples or "snapshots" to one-half of the FFT size (i.e., 4096, sagely applying Nyquist's sampling theorem - possibly in reverse!), but I do not know the veracity of this thought.  I earlier stated that my experience has taught me to "dabble" (experiment) with the "Reduce by" setting (in decibels) to find out what works.  Having said that, one setting is NOT applicable to all situations; therefore, a little "magic" has to be invoked at times.

Thanks for your kind response and patience with this aging ol' fool...

Sounds to me like you've got a good handle on the theory, RNeff.  So just keep playing with different settings and use your ears.  A lot!  Very helpful is the "Keep Noise Only" function that allows you to hear whether the settings and noise profile are doing their job well.  If they are, you'll hear very little desired signal.  Of course, the less of that, the better.

I agree that every job needs its own custom tweaks.  The varying nature of the signals, not to mention the sources, demand that.  However, I tend to use some wacky settings as a starting point, like:
Snapshots = 4000
Noise Reduction Level = 72-78
FFT Size = 8K for most, 6K for percussive sounds, and 12-16K for smooth, sustained sounds
Reduce by = 40dB, because with the Noise Reduction Level set to 74 or so, you're getting WAY less than 40 dB
Precision factor = 13, just for whatever
Smoothing amount = 1
Transition width = 0
Spectral Decay Rate = 0 or at the most 1.  This is my least favorite feature in NR because it creates, to my ear, dreadful noise pumping and dramatically decreases the good that the other settings accomplish.

For what it's worth!

OK 'MusicConductor' - some more numbers to 'play' with!  Every little bit helps and these parameters I have not yet tried.  Thanks for the input!
R Neff

[Voodoo First Law of Noise Reduction:]

OK, Steve!  Well, let me see.  When all else fails, try following instructions...  Back in the "good ol' days" when Steve Fazio was lurking around the old Synt forum, he once said, "Lower FFT sizes of 2048 and below have better response times (I interpreted that as computer execution time) but poor frequency resolution (which makes sense).  Higher FFT sizes of 12000 or more (!) have very good frequency characterization of the analyzed signal but may introduce "whooshing" and reverb-like sounds." 

I'm afraid it has nothing to do with computer execution time at all - it's all to do with the effects of the FFT window size, and how it responds to the time spectra of different signals. (I would have elaborated a bit more in the last post, but I was in Russia at the time, and it turns out that data transmission time the way I was doing it is expensive from there, even though it works quite well...)

Anyway, the long and short of it is that larger FFT window sizes (ie, smaller numbers) tend to work better with LF noise removal, and shorter ones with HF noise, and it's generally the latter that predominates. Some people have found (and it depends rather on the material) that doing more than one pass at NR using less actual reduction, but different FFT sizes each time, can produce a less artefact-free result, although I have to say that on most of the things I  use it for, the highest settings available tend to produce the best results - primarily because I don't generally encounter a lot of LF noise. If you try to use a large window for all noise, then the artefact level goes up a lot, whatever you set the slider at - the relatively large sample is never going to match the small noise sizes, whereas it may well work better with the larger LF ones. Unfortunately the warbling sound is going to be a lot more dominant at higher frequencies though, which is why generally higher FFT numbers are better.

The other thing to note is that Steve F. was talking about earlier implementations of NR; since he wrote that the entire algorithm has been rewritten (twice, I think), and the general artefact level is lower anyway - so the higher settings are generally better now, regardless. Having compared the results you can now get from Audition's NR to expensive systems like CEDAR, on balance I think that I prefer Audition - the CEDAR results always sound 'weird' somehow... although to get AA NR to be that effective, you really have to make quite an effort.

If you look in the archive, you'll find several references to the ultimate self-fulfilling prophecy about NR - and that is that it works best on signals that are already pretty quiet in the first place!

It was actually posted originally like this:

Voodoo First Law of Noise Reduction:
"Noise reduction works best on signals with..................... very little noise"!

But whatever you've got, it's always worth experimenting first.

Thanks again, Steve!  Who knows, one day I may actually get a handle on this business of audio restoration - including effective noise reduction...