1.2.6 noise removal superior to 1.3.6 ??

I have transferred this as a Feature Request to the Pending Feature Requests page on the Audacity, where it will be triaged bty the Audacity developers. I have retained it on the forum, but moved it to the Audio Processing section as it contains some useful information for forum readers. Waxcylinder

  1. Windows XP
  2. removing cassette tape background hiss

well I’'ve played and played with the beta version 136 (& 135) sliders and in pure terms of just listening to the result I simply cannot get as good result in 136 or 135 as I do in 126. Versions 136/135 DO a good job in eliminating the hiss in the area BETWEEN songs but as soon as the music starts again for some reason remnant hiss returns for me - the top and bottom sliders in 136 come the closest in eliminating it (“noise removal” & “attack/delay” - bottom one (attack/delay) has to be all the way left) but no doubt for me its still there whereas 126 pretty much gets rid of all of it for me (the single slider usually in the position 2nd from left). Well I’m no techie and don’t know why this is or exactly what the sliders are doing - just reporting as it sounds to me after much experimentation, sliders in many different positions and combinations - maybe its just me. --LC

“Generally” the noise reduction in 1.3.x is better than in 1.2.6 - that is to say, if you are removing low level noise such as tape hiss, and make a good noise profile, and use optimum setting, then with the majority of source material, subjectively better results can be obtained using the noise reduction in Audacity 1.3.x

However, with some types of noise, and with some types of source materials, Audacity 1.2.6 may do a subjectively better job. This is partly a matter of personal taste/preference, and partly due to the type of sound/noise you are working with.

The noise reduction in 1.3.x uses a different algorithm to that in 1.2.6, and in nearly all cases I find that I get better results with 1.3.x, though I have on occasion found 1.2.6 to be better. For this reason, I keep both versions installed so that I have the choice.

“much experimentation” is a good approach. Only you will know which you like best.

(I also have “Gnome Wave Cleaner”, which is a dedicated audio restoration tool for Linux which I think does a better job at noise reduction than either Audacity version, though it is a little more complicated to set up).

thanks for reply (I’m the guy who started this thread - 126 noise removal better for me) - maybe its my old cassettes, maybe my (not new) equipment, who knows? - yes for that reason i also have to hang onto version 126 - maybe hopefully final stable version 1.4 will have the old 126 noise removal option built into it so I don’t have to change back and forth!
:smiley:
(the (discontinued) Microsoft “Plus! Analog Recorder” program also did a very nice noise removal quality job - I used it for a long time but it left you no archival master track to work with in the future if wanted - not nearly as versatile as audacity but still a very nice job for the very specific task of LP record and cassette tape digital conversion)

I got a note once from one of the program daddies and he said that if you turned off all the extra adjustments in 1.3.x, you got the same noise removal as in 1.2.x. One was built above the other.

I have never found the earlier noise removal tool to be useful at all. If you have the one job that it’s good at, then you win.

So you already have both.

Koz

I got a note once from one of the program daddies and he said that if you turned off all the extra adjustments in 1.3.x, you got the same noise removal as in 1.2.x. One was built above the other.

tried all that - somehow still not the same for me or as good as 126 - its very strange - 135/6 does a perfect job on the (hissy) ‘silent’ spaces BETWEEN the songs but when the sound kicks in again I pick up a renewed kind of intermittent background hiss no matter how I manipulate the sliders - seems very odd.

I have never found the earlier noise removal tool to be useful at all. If you have the one job that it’s good at, then you win.

I’m starting to feel like the “special” child…
:astonished:
:slight_smile:

So you already have both.

true thank goodness - I’ll keep playing with 136 with future source cassettes - maybe things will change (like maybe my ears)
:slight_smile:

also - in final stable 1.4, in the noise removal instructions, wonder if would be possible to give three small visual ‘soundwave’ pics of what each of the 3 sliders are ‘doing’ - would help out the illiterati like me.
:slight_smile:

You’ve got my curiosity going koz.
Is it possible for the user to “turn off all the extra adjustments in 1.3.x” ?

The following WAV files are each about 280kb.

I’ve been experimenting with this, admittedly artificial, test.
http://easyspacepro.com/audacity/examples/noise_removal/noise.wav

Attempting to remove the background white noise with Audacity 1.2.6
http://easyspacepro.com/audacity/examples/noise_removal/noise-remove-126.wav

Attempting to remove the white noise with Audacity 1.3.6
http://easyspacepro.com/audacity/examples/noise_removal/noise-remove-136.wav

On this example, no matter what settings I try, I just can’t get Audacity 1.3.6 to do as good a job as Audacity 1.2.6
I’ve tried with Audacity 1.3.3 and 1.3.4 on Windows and Audacity 1.3.6 on Linux, and all versions of 1.3.x yield the exact same results.

If it is possible to reproduce the 1.2.6 effect in 1.3.x, how the heck do you do it?

AHA - SO I’M NOT THE ONLY ONE!!
:smiley:

No your not. As I said in my first post, the noise reduction in Audacity 1.3.x is usually better, but not always.

I’ve found some clarification on what koz was saying about them both being based on the same code.
http://audacityteam.org/wiki/index.php?title=Noise_Removal#Improved_Noise_Removal_in_1.3.3

In the new Noise Removal in Audacity 1.3.3, there are now two additional sliders to give more control over the removal algorithm - one controlling the smoothness in the time domain via an attack/delay control, and the other controlling smoothness in the frequency domain. If you set both of these to zero, you’ll get something > very similar > to the old noise removal algorithm in Audacity 1.2.x

Similar, but not the same. I’ve not analysed the code (I wouldn’t know how to), but it seems to me that Audacity has fixed (but not zero) values for the smoothness parameters. I generally find that values of 150Hz and an Attack/Decay time of about 0.14 seconds to be about the optimum, though this can vary with different types of audio.

The crucial bit which explains the difficulty of reproducing the 1.2.6 type of effect with 1.3.x:

The main slider controlling the amount of noise reduction works differently now: instead of adjusting the threshold, it adjusts the gain applied to the noisy part of the signal - i.e. the amount by which the noise should be reduced in volume.

So if we want to have the same functionality of the 1.2.6 effect in 1.3.x, we need the “Threshold” slider.

Returning to the issue of my test samples - these are not very realistic test conditions. I deliberately created a sample that Audacity 1.2.6 would handle better, just to illustrate the point. The level of noise in that sample (around -40dB) is much louder than you should ever have in a recording, and if you do have that level of noise, you should go back and record it again.

Although it is not very evident in the listening test, the 1.2.6 noise reduction has done quite a lot of damage to the signal that I was trying to “restore” - particularly to the initial attack of the sound, which if you load the samples into Audacity you can see have been noticeably mangled. On the other hand, although there is still noise present in the 1.3.x version, (and it is quite an annoying “swooshing” sound), the audio signal that I was trying to “restore” is much more intact.

Having done some more experimentation, I have found a workaround that allows similar (in fact better) results using Audacity 1.3.x for this test sample;
The method is to make a duplicate of the “noise”, and amplify it before making the noise profile. This in effect changes the noise threshold value that we would be setting in Audacity 1.2.6

In this sample, I duplicated the “noise” and amplified it by 20dB, then used that for making the noise profile.
I then applied the noise reduction to the original sample with the following settings:

Noise reduction: 32 dB
This was sufficient to make the white noise virtually inaudible, but as low as possible to avoid damaging the original too much.

Frequency smoothing: 120 Hz
This was sufficient to prevent the resulting sound from “wobbling”, or gaining a hollow metallic sound. Raising this level too high would again start to damage the sound that want to keep, making it sound muffled.

Attack/Decay: 0.35 seconds (quite a lot higher than I usually use it, but necessary because the noise level is so high)
Setting this too low tended to clip off the natural decay of the note. Setting it too high and we get that “swooshing” sound as the noise fades in and out.

The result can be heard here: http://easyspacepro.com/audacity/examples/noise_removal/noise-removal-better-136.wav
I think that you will agree that this is better than either of the previous attempts.

The upshot of all this, is that it would be good to have the threshold slider back, but in addition to the refinements that we currently have in 1.3.x

If the developers think that this makes the effect too complicated, perhaps with the categorisation of the effects menu, there could be two versions, a simple, and an advanced interface.

This seems to have developed into a “feature request”, so I’m moving it to that section of the forum.
What do you think?

yes I’ve read that section several times - technically mostly lost on me. Your other comments above (edited here) are very useful and educational and I will continue to try and work with them.

The upshot of all this, is that it would be good to have the threshold slider back, but in addition to the refinements that we currently have in 1.3.x

If the developers think that this makes the effect too complicated, perhaps with the categorization of the effects menu, there could be two versions, a simple, and an advanced interface.

This seems to have developed into a “feature request”, so I’m moving it to that section of the forum.
What do you think?

For me right now that would be a very good thing! (having a ‘simple’ 1.2.6 version option in final 1.4) - steve thank you very much for your efforts in this area!

So you would like to add your vote for this feature request?

Namely; Improving the Noise Removal effect in Audacity by adding the “threshold” control that was used in Audacity 1.2.x and offering both a full version (with the additional slider) and a simplified interface that used a single slider for more/less noise reduction.

I think that in the simplified interface, could combine both the “threshold”, and the “amount by which the noise should be reduced” in a single slider, and fixed values for “attack/decay” and “smoothing” (probably fixed at the current default values that are used in Audacity 1.3.x). At low amounts of noise removal, the effect would be more like the 1.3.x effect, then as the slider was increased it would become more like the 1.2.x effect.

yes absolutely - is there ‘someplace’ to formally ‘go vote’ on that?

Namely; Improving the Noise Removal effect in Audacity by adding the “threshold” control that was used in Audacity 1.2.x and offering both a full version (with the additional slider) and a simplified interface that used a single slider for more/less noise reduction.

I think that in the simplified interface, could combine both the “threshold”, and the “amount by which the noise should be reduced” in a single slider, and fixed values for “attack/decay” and “smoothing” (probably fixed at the current default values that are used in Audacity 1.3.x). At low amounts of noise removal, the effect would be more like the 1.3.x effect, then as the slider was increased it would become more like the 1.2.x effect.

Thanks for your efforts in testing, Steve. I honestly don’t think two alternative Noise Removal effects is sellable to the developers. Better to keep the time and frequency smoothing controls, rather than remove them and set a default smoothing.

I think part of the problem with those two controls is the “jargon”. However the draft Beta manual says more simply what these do and when it might be best to fiddle with their settings:
http://www.audacityteam.org/manual/index.php?title=Noise_Removal

There is also a link on that page which explains in a more technical way about Noise Removal and the “tinkling bells” phenomenon, which is much less of a problem in the 1.3.x removal.

Longer term, presets for particular types of noise such as “hiss”, “hum”, “vinyl roar” and similar might be worth thinking about.

Dominic studied a fair number of “real life” noise/music samples in designing the 1.3 Removal. Clearly Noise Removal doesn’t work well with Steve’s test file, but it’s test bed white noise specifically designed to have equal energy across the frequency spectrum, so it’s always going to be similar to the non-noise. LCoop, could you post a link to a real world test file where in 1.3 you could not remove the noise from the “noise and music combined”? It only need be a few seconds with some noise on its own, then music with the noise, to give us something to work on.

The volume of the noise always did affect the result in 1.2.6 too, whatever you set the threshold slider as, hence why you sometimes got a better result by reducing the amplification of the Noise Profile. I’m not sure how sensible combining a threshold/dB reduction slider is, or how much an extra Threshold slider will help with 1.3 Removal, but I’m not an expert here. Dominic is expert, so I will try and get him interested in this, but I can’t guarantee anything as he is very busy with non-Audacity work now.

As for adding to Feature Requests on our Wiki, I’ve already got this held for review on our Pending Feature Requests page, so at present there is no point adding it as a Feature Request.


Gale

Hi - I’m the author of the noise removal code. I would be interested in trying to improve the code in 1.3.x so that it is always better than the old 1.2.x code, not just sometimes.

To help me out, please send me noisy samples! Ideally I’d like a clip of at least 10 seconds but no more than 1 minute, with at least 3 - 4 seconds of silence (including the background noise). I listened to a bunch of my own recordings, plus a few others that people sent me, while tuning the algorithm, but the more examples I have, the better job I could do.

Please email the noisy samples to dominic at audacityteam dot org.

I’m most interested in background noise / hiss, like the kind that the Noise Removal effect is good for - but I’m also interested in pops and crackles, though that’s a different type of noise.

Thanks!

  • Dominic

I hope Audacity will “do something” about
the noise removal effect. I have not been able to remove
noise from any of my VHS captured soundtracks
with the “new” Audacity noise removal.
The “new” Audacity noise removal seems to remove the noise completely
from the noise sample but not at all from the soundtrack.
The “old” Audacity noise removal might have sometimes
left squeaky residue/side-effects, but at least it would remove the noise.

Enabling the ReaFir would be the best solution.
http://reaper.fm/files/reaplugs.zip

ReaFir would not have to be “real-time” if Audacity would at least
apply the ReaFir settings without having to draw the line in precise mode.

It was explained to me firmly that the early noise reduction is not a stripped down version of the modern one. They used different techniques leading almost effortlessly to people who swear at one and by the other…and they’re different.

There are vinyl users who like the early one and consider the new one a waste of time. I’m the exact opposite. I’ve never gotten the early one to do anything useful.

It seems to me, and history is firmly on my side, that people want the practiced, excellent, top quality results of the second one with the simplicity of the first.

“What is this ‘smoothing’ thing? Just give me a slider with more and less correction.” The bumper sticker version of a complicated issue.

Because they do act differently, I can make a case for including both. The “red” corrector and the “blue” corrector. The Final Cut Pro and iMovie versions. Or Photoshop and Microsoft Paint. Pick your two appropriate versions of the same thing.

You could even fold the two versions into the same operator’s panel. [Click Here For Expert Tools].

My only change to the modern one is a one-click that lets me hear what noise reduction took out. Yes, I know there is a fifteen minute series of tools I can apply to do that, but I’m too close to retirement.

Koz

Quite often I will boost my “noise sample” by around 6 dB prior to capturing the noise profile. I find that with the type of material that I work with this gives a happy medium between the old and new style algorithms and usually gives better results than the straight application of either.

There’s a few effects that could benefit from that approach, but it requires a whole new level of complexity in the interface, so I doubt we’ll be seeing such a thing very soon. On the other hand, if/when “Effect Categories” is reintroduced (with the necessary refinements to make it more usable), there could perhaps be a differentiation between simple/expert tools with the option to hide one or other.

It’s open source, so you’re free to try your hand at improving it.

Marco Diego Mesquita has provided a possible patch that adds a “Sensitivity” slider to the 1.3 Noise Removal. Moving this slider rightwards to a positive dB value raises the threshold and so cuts more signal. Sliding rightwards all the way to 20 dB gives a similar result in Steve’s “Risset-Drum-with-noise” sample to using the current 1.3 Noise Removal and amplifying the noise profile by 20 dB.

I got a good result at sensitivity + 20 dB ; noise reduction 30 dB; Smoothing 150 Hz; Attack/Decay 0.15s. The drum is very distorted after removal until you move the threshold up to 20 dB, as it was in 1.3.12 - this seems to make more difference to the quality of the drum than the other settings. Even at those settings I still feel the drum sounds slightly more “whooshy” and distorted after removal than in 1.2.6, but the initial attack is cleaner, there is no click at the end, and as much or more noise has been removed with a decent result.

There is also a new button to “Leave only noise” which isolates the noise that would have been removed if you had run the effect normally.

Those on Windows can try it out in the current 1.3.13 alpha. This isn’t necessarily how it will finally be committed, if it is.



Gale

While testing Marco’s original patch I noticed that there is a bug in the current Noise Removal effect.
The “Attack/decay time” slider does not work.

Irrespective of the slider setting, the attack/decay time is stuck at 0.00

This is the main reason why the effect sounds so “wooshy”.

If you try Audacity 1.3.6 or earlier you can reduce the “wooshyness” by increasing the Attack/Decay time.

The effect was rewritten around about Audacity 1.3.7 and from that point onward the “Attack/decay time” control has been broken.

Aren’t we doing a lot of the work a computer is supposed to be doing?

  1. Select a noise profile.
  2. Set the amount of distortion you’re willing to accept in the final show.
  3. Go.

The software tries all the possible combinations of all the tools and variables and selects the best parameters that achieve the desired work…while I’m making a cuppa.

If the noise didn’t go far enough down, you get to choose a better profile or content yourself with increased distortion. Anyone trying to tune individual delay, depth, and smoothing values over hours of iterations is wasting their time.

Koz