Problems with the new Noise Reduction replacing Noise Remova

[Gale Andrews]

On the old tool of Noise Removal the background is now removed and no effect on the rest of the audio. Using the new Noise Reduction tool you have the metallic sound left on the rest of the sound.

Your "post Noise Reduction" sample sounds slightly worse than what I get, but I can certainly see and hear what you mean. The difference is most likely to be due to you selecting a different section for the noise profile from what I used. I used from 4.127 seconds to 4.871 seconds.

This is the same issue that I raised prior to 2.1.0 being released.

You mean the issue is that there are artifacts in the audio we want to keep? Or also that the new effect is more sensitive to the exact profile chosen?

Do you still find "Mine max smoothing" objectionable?

It isn't much use if what you want to keep is very quiet music and the artifacts are almost as loud as the music.


Gale

[kozikowski]

More real life.

The first Noise Reduction panel (attach) was posted by macmex and he commented that it didn't look like mine — second attach. He was trying to duplicate my settings so to get the same results.

I believe we're both in 2.1.0

Koz

[Paul L]

I also want to remind us that nobody has linked to examples of the OLD effect with smoothing turned off.

Here you are:

NoiseRemoval-no-smoothing.flac

Noise Reduction: 24
Sensitivity: 0
Smoothing: 0
Attack/decay: 0.15

Noise profile same as my other samples.

Now I think the metallic "chiminess" is quite obvious in this one. Sensitivity or smoothing can be increased to mitigate this problem.

I strongly suggest taking a spectrogram view to all who follow this discussion. Learn to SEE the visible signatures that "metallic" sounds make.

[Paul L]

More real life.

The first Noise Reduction panel (attach) was posted by macmex and he commented that it didn't look like mine — second attach. He was trying to duplicate my settings so to get the same results.

I believe we're both in 2.1.0

Koz

Impossible. The first is clearly 2.0.6 or earlier.

[Paul L]

On the old tool of Noise Removal the background is now removed and no effect on the rest of the audio. Using the new Noise Reduction tool you have the metallic sound left on the rest of the sound.

Your "post Noise Reduction" sample sounds slightly worse than what I get, but I can certainly see and hear what you mean. The difference is most likely to be due to you selecting a different section for the noise profile from what I used. I used from 4.127 seconds to 4.871 seconds.

This is the same issue that I raised prior to 2.1.0 being released.

You mean the issue is that there are artifacts in the audio we want to keep? Or also that the new effect is more sensitive to the exact profile chosen?

I think you refer to a theme of my comments here? The new effect is much better at really reducing the background noise by the nominal slider amount, even when the noise profile selection is just 1/2 second. So a lesser setting for the first slider may be sufficient than with the old effect.

Do you still find "Mine max smoothing" objectionable?

It isn't much use if what you want to keep is very quiet music and the artifacts are almost as loud as the music.

Gale

I don't think that describes Darren's example. I think what you describe is an inherently difficult case where you would experiment with less sensitivity and more frequency smoothing to get the best result you can.

[steve]

I strongly suggest taking a spectrogram view to all who follow this discussion. Learn to SEE the visible signatures that "metallic" sounds make.

Why should anyone care what it "looks" like? It's audio - I care about what it "sounds" like.
Ref: "Equation" by Aphex Twin

[Paul L]

I strongly suggest taking a spectrogram view to all who follow this discussion. Learn to SEE the visible signatures that "metallic" sounds make.

Why should anyone care what it "looks" like? It's audio - I care about what it "sounds" like.
Ref: "Equation" by Aphex Twin

I believe it can aid understanding of what the controls of the effect do.

If you are accustomed to spectrograms, I can tell you that the tinkly artifacts in the quiet parts appear as spots on the spectrogram, Sensitivity increase causes fewer spots, while Frequency Smoothing blurs those spots as post-processing while not really reducing their number. Making fewer spots on the spectrogram in the pauses is one of the great improvements in the new effect, and I believe that is superior to reliance on high frequency smoothing just to disguise them, as was the practice with the old effect.

Yet we still need some frequency smoothing.

[Gale Andrews]

On the old tool of Noise Removal the background is now removed and no effect on the rest of the audio. Using the new Noise Reduction tool you have the metallic sound left on the rest of the sound.

Your "post Noise Reduction" sample sounds slightly worse than what I get, but I can certainly see and hear what you mean. The difference is most likely to be due to you selecting a different section for the noise profile from what I used. I used from 4.127 seconds to 4.871 seconds.

This is the same issue that I raised prior to 2.1.0 being released.

You mean the issue is that there are artifacts in the audio we want to keep? Or also that the new effect is more sensitive to the exact profile chosen?

I think you refer to a theme of my comments here? The new effect is much better at really reducing the background noise by the nominal slider amount, even when the noise profile selection is just 1/2 second. So a lesser setting for the first slider may be sufficient than with the old effect.

No - I was asking if the issue Steve raised prior to 2.1.0 was the artifacts in the processed sound, or if it was also that (whatever the length of the profile) the exact profile taken now affected the result more than in 2.0.6.

Do you still find "Mine max smoothing" objectionable?

It isn't much use if what you want to keep is very quiet music and the artifacts are almost as loud as the music.

I don't think that describes Darren's example. I think what you describe is an inherently difficult case where you would experiment with less sensitivity and more frequency smoothing to get the best result you can.

But I can still get a better result for that "difficult case" (no significant artifacts) in Goldwave with much less twiddling.

I know you need samples, but the samples I tested before with Marco's Spectral Subtraction or Jérôme's Noise Coring produced a subjectively excellent result in their different ways, with almost no artifacts.


Gale

[Paul L]

the exact profile taken now affected the result more than in 2.0.6.


Gale

Do you think that is the case in the new effect, or are you only asking if Steve thinks so, and what exactly do you mean by that?

With the old effect, as we saw with the white noise examples, longer profiles of uniform noise resulted in more reduction of the noise. Very long profiles were needed in fact to get reduction that approached the nominal slider settings.

Whereas now, short noise profiles are sufficient for good noise reduction, and diminishing returns to longer profiles set in very early.

If that is what you mean -- that short profiles give greater noise reductions than before -- then I say that is a good thing.

[Gale Andrews]

the exact profile taken now affected the result more than in 2.0.6.


Gale

Do you think that is the case in the new effect, or are you only asking if Steve thinks so, and what exactly do you mean by that?

To be clear, I never use Noise Reduction or Noise Removal except to test it casually because it produces too many artifacts for my esoteric low signal cases. So I don't know if taking audibly similar profiles at slightly different same-length selections affects the result more than in 2.0.6. I was asking Steve if he meant that.

With the old effect, as we saw with the white noise examples, longer profiles of uniform noise resulted in more reduction of the noise. Very long profiles were needed in fact to get reduction that approached the nominal slider settings.

Whereas now, short noise profiles are sufficient for good noise reduction, and diminishing returns to longer profiles set in very early.

If that is what you mean -- that short profiles give greater noise reductions than before -- then I say that is a good thing.

That is not what I meant - see above.

It is good that Audacity works better with short selections.

But "diminishing returns to longer profiles" sounds like it could be another reason users accustomed to capturing a long profile in Noise Removal could not be getting as good an overall result in Noise Reduction as Noise Removal.

Or do you think the same long profile would still give better reduction in Noise Reduction? Note the difference between "better reduction" and "overall result".

The Noise Reduction - Audacity Manual page says

A minimum of 2048 samples (0.05 sec. at 44100 Hz sample rate) is needed, below which an error will be shown. A longer profile is better

Can the sample now be too long?

Gale