Page 1 of 2

Noise Reduction Trick -- UNDO

Posted: Sun Apr 10, 2011 4:30 pm
by kozikowski
We understand clearly the value in Effect > Amplify a noise segment before capturing a Noise Reduction profile. Then UNDO your way back to normal before you apply the tool.

Does UNDO undo the profile capture?

Are you sure?

Koz

Re: Noise Reduction Trick -- UNDO

Posted: Sun Apr 10, 2011 6:55 pm
by steve
"Undo" does not remove the captured profile (yes I'm sure) ;)
In Audacity 1.3.13 the "Effect > Amplify a noise segment before capturing a Noise Reduction profile" is not necessary. The new "Sensitivity" slider does that.
Sensitivity = 0 is the same as no amplification of the noise sample.
Sensitivity greater than 0 is the same as positive amplification of the noise sample.
Sensitivity less than 0 is the same as negative amplification of the noise sample.

Personally I've found no benefit to negative values, but other alpha-testers have reported that in some cases it can be advantageous.

Re: Noise Reduction Trick -- UNDO

Posted: Mon Apr 11, 2011 3:06 pm
by kozikowski
Which setting is the same as Effect > Amplify (Default)?

Koz

Re: Noise Reduction Trick -- UNDO

Posted: Mon Apr 11, 2011 4:27 pm
by steve
kozikowski wrote:Which setting is the same as Effect > Amplify (Default)?
Not sure what you mean Koz.
Do you mean "What Sensitivity slider setting is equivalent to pre-amplifying the noise sample to 0 dB"?

Re: Noise Reduction Trick -- UNDO

Posted: Mon Apr 11, 2011 5:10 pm
by kozikowski
Yes. Like the slider is labeled Normal to Auto/Max.

Does the tool call Effect > Amplify, or call the code?

You say the Developers found a reason to include negative numbers. It always rings bells when Developers/Engineers do that. That means one Developer on a particularly difficult Saturday Night found one (1) instance where a slightly negative number made a barely measurable improvement.

It's also of concern that a slider was designed. That means an already magical series of adjustments just got even more complicated.

Quick. The noise reduction is bubbly. Which control is wrong -- include the new one in your consideration.

Koz

Re: Noise Reduction Trick -- UNDO

Posted: Mon Apr 11, 2011 5:13 pm
by kozikowski
I'm remembering from the dark recesses of my head that Cool Edit had controls on the Profile Capture step. I need to look.

Koz

Re: Noise Reduction Trick -- UNDO

Posted: Mon Apr 11, 2011 5:25 pm
by billw58
kozikowski wrote: Quick. The noise reduction is bubbly. Which control is wrong -- include the new one in your consideration.
Depends on what the current settings of all controls are. They could be asking for 30 dB of NR. Or have sensitivity set to +20 dB.

I still maintain that the sensitivity slider (or the trick of amplifying the noise section before sampling it) is a kludge.

See: http://forum.audacityteam.org/viewtopic ... 28&t=39529

-- Bill

Re: Noise Reduction Trick -- UNDO

Posted: Mon Apr 11, 2011 5:40 pm
by steve
kozikowski wrote:I'm remembering from the dark recesses of my head that Cool Edit had controls on the Profile Capture step. I need to look.
The "Sensitivity" slider in the new Audacity version of Noise Removal is similar to moving the yellow line up/down on the graph in Cool Edit Pro.

My vote was for the slider to have a range of 0 to +20 dB.
On all of my (many) tests, best results were with Sensitivity somewhere in the range of 0 to 16, and more usually in the range 0 to 10.
Gale reported that some (non-developer) testers reported that negative values were sometimes advantageous. My own tests suggest that any 'benefit' to using negative values for the Sensitivity can be equally (or better) achieved by reducing the "Noise Reduction (dB)" slider and leaving the Sensitivity slider at zero.
kozikowski wrote:Does the tool call Effect > Amplify, or call the code?
Neither.
The Sensitivity slider was suggested by Marco Diego Aurél (along with a patch) and was 'inspired' by my tip about pre-amplifying the noise sample. He realised that the effect is just number crunching, so all that is needed to reproduce the effect of pre-amplifying the noise sample is to raise or lower the noise profile, and that's what the Sensitivity slider does.
billw58 wrote:I still maintain that the sensitivity slider (or the trick of amplifying the noise section before sampling it) is a kludge.
See: viewtopic.php?f=28&t=39529
I'd generally agree with that synopsis, but as with the "amplify noise sample trick", it does provide subjectively better results, and Marco is currently looking into further improvements to the Noise Removal effect.

At one point it was suggested that perhaps the "Gnome Wave Cleaner" method could be implemented in Audacity.
Unfortunately Jeff Welty (the author of GWC) is not wanting to port it to Audacity, though he has said that he would in principle be willing to have some involvement in an advisory capacity. For anyone unfamiliar with GWC, it has first class noise reduction.

Re: Noise Reduction Trick -- UNDO

Posted: Mon Apr 11, 2011 6:24 pm
by kozikowski
My only serious objection to these gymnastics is the ability to dig yourself a hole -- to amplify the profile into clipping by accident. That's the one thing that Effect > Amplify would not do without user intervention.

Koz

Re: Noise Reduction Trick -- UNDO

Posted: Mon Apr 11, 2011 6:46 pm
by steve
kozikowski wrote:to amplify the profile into clipping by accident
There's no chance of that happening.

Worst case scenario:
  • The original noise sample is 0 dBFS (OK, so there's no show, but let's go with the figures for a moment)
  • Push the Sensitivity slider up to maximum - the profile may now peak at +20 dB (this is not going to rescue the show, but never mind)
  • The noise profile will not clip - it's just number that (probably) now go beyond +/- 1, but the processing is done in 32-bit float format, which can handle huge numbers, so no problem there.
  • The Noise Removal algorithm will now try to remove the noise, and seeing such a high level noise profile it will attenuate the original audio a lot.
  • The result is likely to be near silence, but there won't be any clipping.