Exploring the new experimental Noise Reduction Effect

Just recently Paul’s code for repairing and improving the Noise Removal Effect has been committed and should be available in the Nightlies (if not now, then very shortly). For the historic discussion leading up to the new code read this thread:
https://forum.audacityteam.org/t/noiseremoval-cpp/34924/1
For an introduction to the new dialog check out the preliminary documentation here:
http://wiki.audacityteam.org/wiki/Proposal_Noise_Removal#Paul_Licameli.27s_Noise_Reduction_effect
Note that this dialog has a lot of controls many of which could disappear eventually after suitable defaults are established by experimentation.

Since the original thread is already 14 pages long and is in a board which may make it hard to discover, I think it makes sense for us to discuss our experimentation in this thread.

The alpha version of Noise Reduction has two sensitivity sliders. Old sensitivity has effect only when the discrimination method choice is “old.” New sensitivity has no negative settings. If you ussd negative sensitivity and had some change, you must have used the old method which is not default.

Can you tell me all the settings you used?

Here’s a good example of a recording which needs a lot of help when it comes to reducing noise:
https://www.youtube.com/watch?v=s3_Ld2msIqQ
This was originally recorded in 1926 by Whispering Jack Smith so is out of copyright. The above link is to YouTube but other versions exist online.

At about 8 seconds there is approximately one second of complete silence from the artist and accompaniment; this seems like a good spot to pick up the noise.

I see there is an inexpensive short book on Kindle by one Frank Dauenhauer that gives a procedure for cleaning up old 78s. It is free with Kindle Unlimited. Do you know this book? Tried its procedure?

Some thoughts on testing procedure…

Find a track which has a real-world example of annoying noise. Remember, there are 2 (at least) types of noise: repetitive short bursts and continuous background. A testing procedure which is good for one type might not work on the other type. If the noise is of the continuous type try to select 20 to 30 seconds which contains both a good example of the signal (e.g. vocals & accompaniment) and also at least one second of pure noise (neither vocal nor accompaniment present); delete all but this section of audio (everything before it and everything after it so that you have a track with only 20 or 30 seconds of audio containing a good sample of the noise and a good representation of the signal). If the noise is of the repetitive-short-bursts variety you might need a longer selection; you want to keep the selection as short as practical (but still long enough to give a good feel for your results) because some of the settings can result in quite time-consuming calculations. Zoom in on the one second of noise and create a region label which encompasses a goodly selection of the noise (.3-.8 seconds). I would also at this time Normalize the entire track. Save this as a Project; if you do not save it on the desktop open the containing folder for easy access. You now have a Project which you may keep opening and testing with – do NOT save any of your test results over this Project, you want to keep it pristine.

Create a shortcut key for this new Noise Reduction effect. Since some of the settings must remain identical between when you “Get Noise Profile” and when you apply that profile I would strongly recommend that you get in the habit of making all of your setting changes before getting the profile; this will eliminate the annoying dialog about settings needing to be the same in Step 1 & Step 2. If at all possible, create a script which will automate applying the profile (this script would send the keyboard shortcut to Audacity which would open the dialog and then send a command which clicks the OK button). If you are on Windows and need help with this step ask me for help.

Open the new Project which you just created and listen to it a bunch of times so that you have the original sound firmly embedded in your short-term memory. Now it’s time to proceed with testing, here are the steps I use:

  1. Using the selection guides, carefully select the noise based on the label you created; you want to do this so that you always select the exact same noise during testing.
  2. Press the shortcut key which will open the Noise Reduction dialog (the OK button will be disabled because there is no profile yet); adjust one (and only one) of the settings.
  3. Click the “Get Noise Profile” button.
  4. Hopefully you have created the script as mentioned above, if so, trigger it, if not, use the shortcut key and either press or click the OK button.

This Effect stores its settings so the next time you open the dialog your previous settings will be recalled; these settings are stored in preferences so they survive exiting Audacity.

The purpose of this effect has always been to reduce noise of a steady sort.

Primary purpose, maybe, but it did used to offer some benefit on repetitive short bursts (quite typical of noise on records).

If that does not work so well in the new effect then that’s a disbenefit IMO, but perhaps we should have an effect targeted at cleaning up records anyway?

I sometimes have to use Goldwave’s “Smoother” effect on the “short bursts” type of noise, but only if the “bursts” are quite varied among themselves, in which case Audacity can do nothing about that noise anyway.


Gale