Auto Profile for Noise Reduction — Proposal

[kozikowski]

-
>>> Note this is a proposal for a future improvement in Noise Removal. It's not a How-To. <<<
-

Sample the whole performance during the Noise Reduction Profile Step and use the predominant sound that doesn't change for the Profile. Or better, take the Profile completely in the background unless you elect to do it manually.

One option could be "Soft Steering."

Default is to sample the whole show, Soft Steering is where you drag-select a large, sloppy portion of the show with the worst noise (or maybe the only part of the show with noise).

Manual is the way it works now. Drag-select the profile sample and you have to hit it exactly right.

Obviously, this is no longer a simple noise reduction branch task. This task has to make very complex decisions. But it's possible. Cellphones, Windows Enhanced Services, and even Macs do it all the time.

Koz

[Gale Andrews]

Sample the whole performance during the Noise Reduction Profile Step and use the predominant sound that doesn't change for the Profile. Or better, take the Profile completely in the background unless you elect to do it manually.

So which one is the request?

Continuous sampling while recording may cause dropouts.

The Noise Coring patch for Noise Removal already gave the possibility to do without the step of grabbing the profile (it could autodetect the noise shape and threshold, or use a profile as now).

Perhaps that "autodetection" point should be added at http://wiki.audacityteam.org/wiki/Propo ... ise_Coring ?


Gale

[kozikowski]

So which one is the request?

Two, added to the classic, legacy version.

— Auto Profile by sampling the whole show.
— Guided Profile by manually selecting a sloppy portion of the show with the majority of the noise.

Auto Profile might be wildly impractical, but no, I don't see it happening during a recording.

The instant you click on Noise Removal, it starts scanning the show? It might pop up a message that warns of slow Audacity response while it's working.

"Extended length shows may take too long to make a Profile. Continue? [Y/N]."
Restrict Auto's time block. It won't go over an hour...

Koz

[Gale Andrews]

So which one is the request?

Two, added to the classic, legacy version.

— Auto Profile by sampling the whole show.

That could be done by the Noise Coring patch.

— Guided Profile by manually selecting a sloppy portion of the show with the majority of the noise.

Please explain how that would work. Are you suggesting that Audacity ignores the selection and scans all the audio anyway, or takes the "Auto Profile" only from within the selection, even if the user selects only a noisy part of the audio-with-noise? They may well do that if we suggest you don't have to be precise with the selection.


Gale

[kozikowski]

I'm pulling that one out of my.... ear.

The "Air Conditioning Noise" is available throughout the entire performance, but there is a segment of the show where the performance is lower volume or does not predominate. Drag-Select this portion of the show — the whole portion — to gently guide Automatic into a more advantageous sample.

It's a given that nobody ever shuts up long enough for classic Noise Removal Profile to work and waiting for Automatic to sample the whole concert might get you too close to retirement.

Koz

[steve]

The "Air Conditioning Noise" is available throughout the entire performance, but there is a segment of the show where the performance is lower volume or does not predominate. Drag-Select this portion of the show — the whole portion — to gently guide Automatic into a more advantageous sample.

I can imagine that working in theory - automatically extract the quietest part(s) of the selection and use that as the noise sample.
I have doubts about how much it will help. It's unlikely to do a better job than manually selecting the noise sample. I get the impression that people generally have more trouble with the second pass than the first - I know that I do.

[kozikowski]

How do you know that trouble with the second pass isn't related to doing a terrible job with the first?
Koz

[steve]

How do you know that trouble with the second pass isn't related to doing a terrible job with the first?

Yes, of course it could be. Without a "good" noise profile the noise reduction will be bad, but you know as well as I that even with a "perfect" noise sample the noise reduction can be quite poor.

I would see more point in polishing the usability if the effect worked better. The proposal as it stands may make the effect easier to use (which would be good), but at best the effectiveness will be no better than now. It would be difficult to test how well the automatic noise profiling works as we would comparing different shades of mediocre.

"Noise coring" that Gale mentioned takes a different approach to the Audacity Noise Removal effect and does not require a noise profile. It is a one pass effect. It currently only exists as a prototype.

[Gale Andrews]

It's a given that nobody ever shuts up long enough for classic Noise Removal Profile to work and

So what is the minimum required length of the profile? Does it depend on sample rate?

On a quick test, a profile of 10 ms at 44100 Hz has no effect on the noise.


Gale

[kozikowski]

Whatever FFT requires to get enough to chew on.
Koz