Removing vocals ONLY, and not other instruments

[Robert J. H.]

I am glad that you're fairly happy with the result.
I've used a lot of things that are actually not yet published.
Firstly, I don't use the Audacity vocal remover because it produces only a mono output and does not support "Isolate Center".
About half a year ago, I've written my own tool for this. You find the newest version of the "2D Stereo Toolkit" in the drop box folder (link above).
Put it in the plug-ins folder and restart Audacity.

If I recollect correctly, I've firstly duplicated the original track (after recording from your source).
I then applied "Isolate Vocals (inverted)" from the action menu of the plug-in mentioned (under effects, 2D Stereo Toolkit) on the copied track.
I've changed the low cut frequency from 120 to 210 Hz to have more bass preserved.
This gives a center channel. You can split this track to mono since the both channels are identical (track drop down menu, split to mono).
Mute one of those or delete it.
That's about all in order to produce an ordinary karaoke version.

In a secondary step, we have to bring back the beat. That's a rather ticklish task.
That's a thing that will go next into the plug-in. Unfortunately, the code is not yet mature and I had therefore to work from the Nyquist prompt (also effects menu). The "Drums" track is simply created by extracting the loudest peak in comparison to its neighbourhood. The ultimate goal is to extract the frequencies that are common to all these peaks, this should give the drums without vocals.
But I don't want to go into detail for now, maybe we can handle the "Drum Separation" in another post. It might be sufficient for now to say that a new created beat would be the best solution for the time being.
I put another plug-in into the karaoke folder, named "Tempo Teller" (appears in the Analyze menu). It returns the estimated beats per minute. It helps to adapt a new beat or to simply boost those times with a kick or snare beat.
However, all this comes later.

By the way, I've not used the any of the secondary sound examples you've linked to. They sound somewhat strange to me, as if they were from a totally other song version, I've no idea how it is made or treated in the first place.

[Gray Fox]

I am glad that you're fairly happy with the result.
I've used a lot of things that are actually not yet published.
Firstly, I don't use the Audacity vocal remover because it produces only a mono output and does not support "Isolate Center".
About half a year ago, I've written my own tool for this. You find the newest version of the "2D Stereo Toolkit" in the drop box folder (link above).
Put it in the plug-ins folder and restart Audacity.

If I recollect correctly, I've firstly duplicated the original track (after recording from your source).
I then applied "Isolate Vocals (inverted)" from the action menu of the plug-in mentioned (under effects, 2D Stereo Toolkit) on the copied track.
I've changed the low cut frequency from 120 to 210 Hz to have more bass preserved.
This gives a center channel. You can split this track to mono since the both channels are identical (track drop down menu, split to mono).
Mute one of those or delete it.
That's about all in order to produce an ordinary karaoke version.

In a secondary step, we have to bring back the beat. That's a rather ticklish task.
That's a thing that will go next into the plug-in. Unfortunately, the code is not yet mature and I had therefore to work from the Nyquist prompt (also effects menu). The "Drums" track is simply created by extracting the loudest peak in comparison to its neighbourhood. The ultimate goal is to extract the frequencies that are common to all these peaks, this should give the drums without vocals.
But I don't want to go into detail for now, maybe we can handle the "Drum Separation" in another post. It might be sufficient for now to say that a new created beat would be the best solution for the time being.
I put another plug-in into the karaoke folder, named "Tempo Teller" (appears in the Analyze menu). It returns the estimated beats per minute. It helps to adapt a new beat or to simply boost those times with a kick or snare beat.
However, all this comes later.

By the way, I've not used the any of the secondary sound examples you've linked to. They sound somewhat strange to me, as if they were from a totally other song version, I've no idea how it is made or treated in the first place.

Hey, thank you so much!!! I've been busy in the past two days with that plugin and tried a few things! I followed your steps and experimented on a few different tracks, I am amazed by the stereo outcome and I managed to mix some tracks where I needed parts of other instrumentals and you barely hear it's inverted is where you normally would with the normal inverting technique. You were right about using a CD quality version, I used a flac file and the main beat was way more present, I'm really happy it keeps most of the main instrumentation and beats but deletes almost all of the vocals at certain points!

What I notice is that the secondary track is close to a filtered acapella, is there a way to use that plugin to mainly keep the voice, or is that impossible?
I'm curious to know how the drum separation goes and how the "Tempo Teller" plugin works!

Once again, thank you so much!

[Robert J. H.]

There's certainly a lot to discover in that plug-in. It is actually a program within a program. That's the main reason why it isn't included in Audacity.
I always try to improve the algorithms. Especially "Remove/Isolate Vocals" should eventually be smarter in order to really concentrate on voice (-like) sounds.
We have essentially those sounds within a song:

• pitched instruments
• voices
• percussion
• noise and reverb artefacts

There are some key points or "features" that can serve to discriminate one kind from the other. However, the boundaries are rather fuzzy. Otherwise, no one could do beat boxing for example.
I do first try to eliminate (regular) percussion beats from the center channel.
It is obvious that the probability for a percussive sound is higher for frames (small parts of the audio) that do fall onto a multiple of the actual tempo.
That's why I've written the tempo teller plug-in. It constructs a table with all magnitudes from 30 to 700 bpm and searches for the highest peak(s).
That's by no means a trivial task because often the peak does not correspond to the felt tempo.
Nevertheless, it is working better from version to version.
I hope that it can soon be integrated into the 2D Stereo Toolkit to cut or boost the percussion.
After that, other instruments will be attacked.