Noise Removal Tests

Steve, can you make a bunch of tests and come up with an optimal value for the FFT window size?

If the problem is AudNR becoming too complex, we can set a “global value” for FFT window size. A value that would be used by every audacity effect that uses FFT. The FFT window size would then be set from a single point in the UI, maybe the preferences window. What do you think about it?

Woohoo! :smiley:
Thanks, Marco.
Since this patch is so simple could you submit it and get is committed? Then I could try it out in the next nightly build.

What is the FFT window size currently set to?

Don’t know about adding a control. I would welcome it, but less experienced users might be turned off, as Steve noted.

– Bill


Thinking back to the days when I used CoolEdit Pro, the current setting is toward the low end.
There’s probably going to be some trade off - possibly on processing time / CPU load.
This calls for some testing. I’ll see what I can do over the next few days.

I’ve tested Noise Removal with FFT set to 8192

Compared with an FFT size of 2048, more noise is removed with less damage to the remaining sound.
The differences are most noticeable with “synthetic” tests (tones over low level noise), but also apply to real world test samples, though the improvement was generally quite subtle with the real world samples that I tested.

The low frequency artefacts mentioned previously are considerably reduced, and in conjunction with a slightly raised “Sensitivity” setting (around +4) are virtually eliminated.

High frequency damage can cause “denoised” audio to sound muffled. This unwanted effect is substantially reduced by using a larger FFT size.

In no tests were the results worse with a larger FFT size.

The down side is that processing time increased by around 20%.

The question is, do people want a relatively subtle improvement at the expense of slower processing?
I’m +1 for improved Noise Removal.

In my books the gold standard is Brian Davies’ DeNoise program. I never use Audacity’s Noise Removal (sic) effect. When Audacity’s NR can produce results comparable to DeNoise then I’ll be happy, and would be fine with increased processing time in exchange for an improved effect.

– Bill

Real world sample here:

Great Steve! You’ve just shown that AudNR is now no longer too far from state of art noise reduction effects.

I have many changes queued for AudNR (you would be very happy to use my “own branch”), but they won’t land in audacity anytime soon because of a “feature freeze” for 2.0. Some of my changes significantly reduce AudNR wooshiness and periodic artifacts (the intermittent sea-shell-like sound). So, I want to find a way so that I can post my changes and interested people can test it. I really need outside opinions and tests.

Do you have an idea on how we can achieve it?

It would be nice if they could be eliminated without raising the sensitivity.

Did you try the noise in the presence of tone test that started this thread? I’d love to know if HF noise is now reduced in the presence of a mid-frequency tone.

What real-world samples have you used? My acid test is a decaying guitar note into noise. A decaying piano note would work as well.

– Bill

It’s not as bad as it used to be.
Unfortunately at least one of the developers think that the current release version of NR is good :confused:
Having used many other Noise Reduction tools, there is IMO, still some way to go, but results so far are very encouraging.
I am personally very keen to see this feature improved as good Noise Reduction would be a very good “selling point” for Audacity, and it would save me the hassle of switching between applications when I have to clean up noisy source material.

As there are several different approaches involved I think it would be very helpful to have an overview page on the Audacity wiki. We could also have a set of test samples on there which would allow fair comparison between different NR methods.

Personally I find wiki formatting highly irritating and totally not user friendly, but in this case I think it would probably be the best format for holding an overview and a review of test results. If we include the posts in this topic, there have already been over 10 pages of posts on this subject, so I think an overview page is essential for keeping track of developments.

Since you (Marco) are most actively involved with developing this feature it would probably be best if you started off the wiki page, but if you’re uncomfortable with posting to the wiki I could get it started off on your behalf.

For quicker and easier access to posting new patches, discussing ideas, and returning test results, I can make this a “sticky” topic on this forum board. This will hopefully encourage other Audacity users to contribute with ideas and test results.

No, not yet, but I’ll try that.

Yes, that would be a good “revealing” test. That’s exactly the kind of thing that I have in mind as part of a set of “standard” test samples.
Do you have a suitable recording with real world noise that you could donate?

Yes Steve, I would like better if you start the wiki page. However, there are some sections that I think are needed:

1 - An Introduction section saying that noise reduction is not an easy task and in this we collect some tests and experiments to check how the effect evolves with changes.

2 - An Experiments section that should contain a list of patches with descriptions of what they do, in which audacity version should be applied, who created, when was created and what improvement is to be expected.

3 - A Results section that would contain a list of noisy samples, their description, the best result achieved with a stock version of audacity (and the settings used) and the best result achieved with an experimental version of audacity (and the settings used).

4 - A conclusion section discussing the results and showing which experiments are more promising.

The beginning of the article should also contain the description of the above four points. That way, people interested in helping will be able to follow a “standard format”.

If we do that, many people will be able to collaborate to improve AudNR, be it by making tests, commenting or posting audio samples.

Not “real world”, but decaying guitar harmonics with white noise added at -40 dB. The last second is just noise.

I’ll go through my vinyl collection. I think I have one that would be suitable. Stay tuned.

– Bill

Here’s the real-world sample. From vinyl, complete with vinyl surface noise and hum.

– Bill

I remember that in that very long thread I started when I first joined this forum I posted a few samples of my tests with the T-Bone USB micplug that were noisy. I can also record a few more samples if you’re interested.

I have started a wiki page here:

I think I have covered most of those points, but I have done so following the standard wiki Proposal format.

I’ll upload, or link to some suitable test samples shortly, but at present Flac files cannot be uploaded to the wiki.
If anyone has good real-world test samples, please post them here.
Test samples should be in Flac or WAV format, and should contain both “noise only” and “noise+signal”.
Maximum file size should be 1 MB.

Maximum file size should be 1 MB.

I still don’t think three seconds is enough for good analysis, particularly if you have to wait for a silent portion in the performance.


1 MB in FLAC format will allow about 10 seconds stereo 16 bit 44.1 kHz.

I’ve been testing the new “Noise Coring” effect by Jérôme M. Berger

Here’s a sample of the effect.
The first sound is the original unprocessed sound. It is followed by a “cleaned” copy using the Noise Removal effect in Audacity 1.3.13, and finally a copy of the original that has been processed with the Noise Coring effect.

On my ears it sounds better than audacity’s noise removal :slight_smile:
The sound is not so muffled, keeping some of the brightness of the original recording.

I agree, and another notable feature is that it does not have any of the metallic/bubbly artefacts that are common with Noise Removal.
The effect is specifically designed to deal with “hiss” type noise and I think that it’s showing great potential.

I’d score it 8-2 to coring. The first half second of the Noise Removal attempt sounds really clean to me (better than coring), but then it deteriorates like Steve describes and gets more muffled (borne out by Plot Spectrum, almost nothing above 6000 Hz in the Noise Removal sample after about half a second). But I get a pulsing artefact at 23 s with coring in the dying away that is more noticeable than in noise removal (it is more evident in speakers than headphones).

I’m a bit concerned with yet another effect to try and get your noise removed whereas most software only has one effect. How many controls does noise coring have? I never tried the patch.