Verdict on the settings for new Noise Reduction?

[Gale Andrews]

Default smoothing will be 6 in future versions.

I am pleased we finally increased default smoothing, but don't you think it looks odd to have a default at the extreme of a slider scale?

Gale

My thoughts too.
The extreme should in any case be higher - 10 or 12.
If I'm correct, 6 is approx. 150 Hz and we used up to 500 Hz in the old effect. This might not be necessary anymore with the partially new algorithm but who knows.
What about a track with 16 kHz sample rate?

Robert

Remember that frequency smoothing counts bands now. The equivalence of 6 to the old 150 Hz assumes a 44100 sample rate.

So what is the equivalence of 6 at for example 96000 Hz, or 8000 Hz or 16000 Hz (which a voice track might be)?

It seems you are still discouraging frequency smoothing above 6 by setting 6 as the maximum. Are you so discouraging?

https://soundcloud.com/blizzkrut/sets/a ... nd-problem which is voice with heavy noise (or a normalized version with DC offset removed) needs more than smoothing of 6 in my opinion to reduce artifacts on the voice in the noise-reduced result.

What settings would you use on that? The Forum topic those samples comes from is http://forum.audacityteam.org/viewtopic ... 46&t=87385.


Gale

If I don't err, the smoothing over 7 bands (since factor 0 = 1 band) for different sample rates are as follows - all in Hertz, except last column:

Code: Select all

Sample Rate	Bandwidth	Smoothing factor
8000	27.34375	37.5875
16000	54.6875	18.29375
22050	75.36621094	13
32000	109.375	8.646875
44100	150.7324219	6
48000	164.0625	5.43125
88200	301.4648438	2.5
96000	328.125	2.215625

The last number shows the factor needed to get the same 150 Hz band width as with 44100/6.

Of course, the window length in seconds changes as well and thus the results are presumably not comparable - I haven't made all the tests needed.

Robert

How did you calculate those numbers?

A "band" means a "bin" with a 2048 sample window. It is the higher sample rates that have narrower bins.

Can you translate that to something that would be useful for the Manual? For a higher sample rate than 44100 Hz, should the user set smoothing higher, other things being equal?


Gale

[Gale Andrews]

It seems you are still discouraging frequency smoothing above 6 by setting 6 as the maximum. Are you so discouraging?

The higher the "smoothing" factor, the more damage is done by the effect.
It may not be so noticeable with speech recordings. but with high quality music recordings the damage is very evident with the new default settings if you listen to the decay of notes.

As a simple demonstration:

1) Generate a few seconds of Pink Noise and Normalize to -46 dB (about -60 dB RMS). This is a reasonable good approximation of low level noise that one might want to remove from a good quality original recording.

2) Import a really good quality music recording that has very low noise - piano music would be a good choice for demonstration purposes.

3) Create a noise sample from the Pink Noise

4a) Apply to the piano music with "Residue" selected. Note how much of the music is present in the residue.
4b) Apply to a mix of the noise and the music - notice how the notes decay much faster than the original.

Yes, I know. I dislike that, and I sometimes miss the ability to control the release.

But the whole thing is a trade off. With higher noise levels, or quieter music, the loss of attack can be subjectively less bad than artifacting.

Are you assuming attack and release must be coupled? That is not so, I made them separate controls.

No, I never liked them being coupled in the old effect. I just meant that sometimes I notice that smoothing seems to lose attack as well as release.

Ah well, that might be remedied by allowing the attack and release sliders to be visible again, (release more so than attack,) which is another debate.
It's a debate that we had about 12 months ago.
A problem adding more controls is the high degree of interdependency between controls.

It is a problem, but more controls could be hidden behind an "Advanced" checkbox, or perhaps separated into another box.

JC opposed the idea of "Advanced" settings and I believe blocked the attack and release sliders because no one had written good enough documentation.

We have an entire effect, albeit disabled by default, called Classic Filters that does not have good enough documentation.

If you provide documentation I will probably support reintroducing (separate) attack and release sliders.

The other side of the coin is - are the time hidden time smoothing values being applied now as optimal as they should be, so that some frequency smoothing can be made without impairing (attack or) release so much?

Or should hidden time smoothing be smart enough to vary with the frequency smoothing setting?

Also do we have to have the interdependency? Could some filter to mask artifacting be applied "over the top" of the effect, like some kind of post-processing?


Gale

I do not understand. Frequency smoothing is indeed already a post-processing which is the last step in the calculations.

As I understand it, Steve argues against time smoothing controls because if they were exposed to the user and user changed them, then the user may have to readjust all the other settings again to get an optimal result, and the whole experience becomes too "fiddly".

If that's the case, is that post-processing operating optimally?


Gale

[waxcylinder]

We have an entire effect, albeit disabled by default, called Classic Filters that does not have good enough documentation.

Mainly because the effect's author couldn't be bothered to engage with the documentation processs and provide some additional input and editorial review

And so it is likely to remain forever that way (as the author has disappeared over the horizon) ...

Peter

[Paul L]

The most problematic noise that I'm finding is low level but uneven noise - in such cases I want the smoothing factor as low as possible (to minimise damage), but that runs the risk of slightly higher noise "breaking through" as tinkles.

That risk is my issue exactly.

But it seems a weakness in the algorithm if it cannot respond to variations in the level of the same noise. And if the noise type changes, it's defeated.


Gale

The old effect would not have been better on that score. Both versions take just one noise sample to train on, and have no notion of adaptiveness to a changing background.

Must we retain the existing algorithm in perpetuity?

Another "Advanced" setting could be a different algorithm.


Gale

I never said we couldn't try to develop some sort of adaptiveness. But it would be a big deal. Would it add much value for many users?

[Gale Andrews]

Both versions take just one noise sample to train on, and have no notion of adaptiveness to a changing background.
[...]

I never said we couldn't try to develop some sort of adaptiveness. But it would be a big deal. Would it add much value for many users?

It could do. Doesn't noise typically show random variation?

One algorithm on the table (noise coring) doesn't even take a noise profile. Does that make it better able to adapt to changing background?

And do you have recommendations about setting frequency smoothing for sample rates other than 44100 Hz?


Gale