Noise reduction experiments (Music)

Ever since I saw a rather interesting video on YouTube of somebody using the noise reduction’s “residue” function to get a profile of a part with no singing and apply that to the song to make an instrumental, I’ve experimented with the noise reduction plugin extensively.

The first thing is: The results will not be usable in any way whatsoever for professional / semi-professional remixes / mashups / what have you. This is merely for curiosity sake, and I personally just do it for fun and nothing more.

For “removing / isolating” an instrument / vocal track, it highly depends on the audio content and the similarities between the parts with the vocals and the parts without. If the song is repetitive and has solo’d instrument portions in the mix then maybe you will have some luck. I actually upmixed a mono performance snippet to a stereo file using noise reduction to get the sources. It’s very computer-y, but it was just an experiment.

Original video (mono):
Mine (stereo):

What I did was I copied the instrumental portions to a new track, made a noise profile of that, and set it to “residue” so I got just the music. Then I opened up another instance of Audacity and did the same procedure, but set it to “reduce” so I got the vocals. I split the music into the 2 channels and added a delay to the right channel. I then made the vocals single-channel mono and added reverb to them (vocal II preset). Oh, and after I mixed it up I generated a new white noise track and used noise reduction to reduce the hissing in the original recording. After that, I mixed the newly created stereo track into an MKV file using MKVToolNix GUI.

Drum tracks can be “isolated” by getting an unrelated drum track (anything with percussive characteristics works, but if possible try to get an isolated drum part in the original mix since it’ll fare better results) and have it make a profile of that and set it to residue. Lower the sensitivity a fair bit (4 or 5. any lower will make it sound hollow).

Bass tracks can be done the same way, but the clarity of the results wildly vary from song to song.

Vocals can sometimes be isolated to the point where it’d be suitable for a non-professional remix by using the Vocal Reduction & Isolation plugin with the preset Isolate Vocals, at strength 50 and the low-cut set to maybe 180-230. Then just reduce the remaining instruments by using noise reduction to capture profiles of the instrumental parts.

In turn, you could make an instrumental track by downmixing this “isolated vocal track” to mono (since it’s dual mono no info is lost) and normalize it to 0 dB, and then just reduce the remaining instruments like I said earlier, and then you can just use the “Invert” effect on it and when the original and the “isolated vocals” play, it’ll cancel out the vocals and leave the drums more intact, all while still being stereo! :slight_smile:

The Party Line is still that Audacity can’t split up a mixed performance into individual instruments, voices and sounds. Performers who post on the forum aren’t interested in experiments. They want something they can use for real musical performances and few of these tricks will do that.

Since, as I think you indicate, Noise Reduction was intended to remove single, stable, non-changing sounds from a performance (hum, buzz, air conditioning, vent fans), the closer your target experiment comes to that ideal, the better it will work. Multiple-octave wide-ranging musical performances need not apply.

There used to be another odd trick you could do with Noise Reduction Profile Selection that would help with echoes, but the effect was minimal and erratic. The Party Line is still we can’t do anything about echoes, either.

Did you notice any differences between ratty MP3 and clean WAV or Music CD? I wonder if the effect would be more valuable if you weren’t fighting MP3 compression damage—although everybody starts with ratty MP3, so it wouldn’t make a lot of real-world difference.


Interesting point. Funny that you mentioned if there were any differences between using an MP3 or a CD rip. I’m planning on re-buying my Amazon MP3 collection (or at least the ones which I need the most) in CD format so I can work on lossless rips. I think that there will be a point in time where software can split up mixed songs with at least some degree of success, with little to no input by the user. Take a look at Hit’n’Mix, for example. That one tries to do full mixes. Their tech is very promising, as their algorithms can separate the backing vocals from the rest of it (the program’s interface is an over-glorified spectrogram, with cartoony and colorful shapes which represent the vocals and other parts and are color-coded). It can also explode chords into the individual pitches, like Melodyne does. However, they’re gonna have to improve upon the “resynthesis” aspect of it, as the results are unusable for much of anything, but I bought an activation key regardless as I like to recreate instrumental tracks to songs and it helps me transcribe the parts and the free version lets you do up to 45 seconds. Oh, and the other promising things that Hi’tn’Mix shows is the algorithms can evidently isolate the parts extremely well. For instance, the bass notes usually have little to no bleed-through from the other elements. The last promising thing is the fact that it’s user-friendly. On paper it sounds straight-forward: Import a song and then select the instrument which you want to remove, and bam. The concept of the program is very neat and it definitely has a market. DJ’s, hip-hop artists, or even corporations who have a mixed recording and they need the parts to be separated for whatever reason would find a use for a program which does this. That, and it’s for both Windows and macOS. The only real competition they have is Audionamix, another company who is trying to take a crack at this whole concept (they do single elements, such as vocals, but I think their professional services can do full mixes).

their professional services can do full mixes

We do note that you can usually do a respectable job, but probably not at Audacity’s cost structure.


Sorry to bump the thread, but here’s an example of what you can do if you know what you’re doing. This is a low quality recording of a song from a Christmas decoration (I put a mic to it while the song played) and I managed to isolate the elements of it. The vocals, brass, and guitar were all able to be “isolated” by using noise reduction and phase inversion.

I think you were part of an earlier thread for ISSE software. That’s supposed to be able to split parts of a performance.


That’s not necessarily what it’s designed for. It was primarily designed to just split up any 2 sources in general, or to reduce noise (which it does a brilliant job at). It’s a pain to work with since it’s so slow with recordings longer than 30 seconds, and it has a few bugs. That, and the spectrogram navigation is horrible.