extract common part of two audios

[mathmax]

Thank you very much for your detailed answer

Assuming that the amplitude of vocals wave is always bigger than the amplitude of the backgrounds waves (which is right most of the time), the signal should be randomly tone down or up (with a probability of 50-50), right?

As a result, the signal shouldn't be globally tone up by stacking the background on the top of the vocals (except when the amplitude of the first is higher than the amplitude of the second)... that is strange... Intuitively, I would have say that a sound become louder by stacking two signals.

I wonder what makes the values positive or negative? I mean, what is the physical meaning of the sign positive or negative?

The sample you attached seems to be a better acapella... but I guess the vocals are altered and as it's intended to be subtracted from other parts of the song (to remove tags), I wonder if the final result will be better or not. When the amplitude of the backings is high, I guess their reduction is more important than distortion of the vocals... so the result should be globally slightly better.

However I wonder if there is a better algorithm to improve the acapella more efficiently.. especially having more than 2 samples, it should be possible to write a script that causes less distortion and more reduction of the backings.
For example, assuming we have 3 samples I know compare a triplet of points. If two have almost the same values and the third is different, there is a good chance that the value of the point in the original vocals is closer to the two identical values. Maybe I could even improve the script working on wider motifs, taking into account more than 1 point... I'm sure we can do a better work having more than two samples. If you have any idea... I provide a third sample:

http://www.mediafire.com/?7hlazcz5ja753sl

We could even work on 6 samples, taking the left and right tracks as each of them are originally stereo.

Please let me know if you have any idea or advice to enhance the acapella. I really need something as clean as possible.

Thanks

max

[steve]

Assuming that the amplitude of vocals wave is always bigger than the amplitude of the backgrounds waves (which is right most of the time), the signal should be randomly tone down or up (with a probability of 50-50), right?

In this picture, the top track is part of a voice recording (zoomed in close)
the bottom track is low level music
the track in the middle is the vocal and the background music combined.

tracks000.png (20.25 KiB) Viewed 1371 times

In the selected section, you can see a dip in the mixed track compared to the vocal track that correlates to the dip in the music track.

Bringing this back to the first two clips that you posted:
In effect what is happening is that the deviations from the common vocal component don't "cancel out", but become a common "noise".
This allows us to use the Noise Removal effect to some good effect.

min-com-and-noise-removal.wav

min-common effect then apply noise removal

(294.36 KiB) Downloaded 70 times

BTW, this is the result from ExtraBoy using Preset "Isolation HR"

extraboy.wav

(149.4 KiB) Downloaded 80 times

[mathmax]

Thank you so much for your answer and the samples you posted.

Indeed I got almost the same result with ExtraBoy... which is too much distorted.

The result after the Noise Removal effect is pretty nice although the DJ Stolen voice is a bit altered. I'll have to check if it allows me to make a cleaner tags deletion on the rest of the song Thank you so much for the time you spent on it.

Did you download the third sample? What do you think of using more than two samples to make a cleaner acapella? Do you think of some algorithms that could still improve the acapella before using the Noise Removal effect? Maybe this would allow to use it with a lower strength and reduce the vocals' distortions...

[steve]

By the time it's back in a mix with something else you'll hardly notice the old backing:

tom-and-DJ.flac

(631.89 KiB) Downloaded 60 times

[mathmax]

It seems that you changed something (the speed?) in the sample "min-com-and-noise-removal.wav" because I can't align it with the original samples...
So unfortunately I can't try to invert and substract the tags from the full song like I did with the first samples. Here is what I get using my samples without the script and the noise remover. I need to know if I get a better result using "min-com-and-noise-removal.wav"


original version
http://www.mediafire.com/?mql0kt099bsxfe3

corrected version
http://www.mediafire.com/?ggupemd8pe21v99

[steve]

So unfortunately I can't try to invert and substract the tags from the full song like I did with the first samples.

You've just changed the rules of the game.
To quote your second post "I don't want to remove them. On the contrary I would like to keep only them... and get rid of the backgrounds."

[mathmax]

no.. sorry for misunderstanding.

First step: create acapella of "DJ Stolen" as clean as possible from several samples
Second step: use the acapella to remove the tags in the song by using the invert method.

In other words I need to preserve the DJ vocals as much as possible in the first step to be able to remove them as much as possible in the second step.

[steve]

First step: create acapella of "DJ Stolen" as clean as possible from several samples

I see, but that's not what you said. The problem with using fancy tools like "Noise Removal" is that it will change the vocal wave - going back to my first post - for cancellation to work the waves must be identical.

For removing the vocals, the method that you used in your last sample works fine - you just need to de-noise it a bit where the first DJ vocal was removed.

de-noise-first_correction.flac

(1.31 MiB) Downloaded 111 times

[mathmax]

Ok, I thought so.. it's indeed more logical and efficient to apply the de-noise tool after the DJ removal.
But still, I think we could improve the final result by making a better acapella first.
Do you think that writing an algorithm based on three or more samples (like I mentioned a few posts above) could help to recover a clean acapella?
Or maybe we could write an script that take the best parts of each samples and reassemble them to constitute a better acapella.

Thank you very much for your help on this work.. Thousands of people are waiting for this rare rare song to be cleaned

[steve]

Do you think that writing an algorithm based on three or more samples (like I mentioned a few posts above) could help to recover a clean acapella?

No I don't think it would help.

Or maybe we could write an script that take the best parts of each samples and reassemble them to constitute a better acapella.

The problem is that although we may be able to "hear" what the best parts are, human hearing is incredibly sophisticated, and trying to emulate that process in software is extremely complex. If that were not enough, the processing must not change the voice part at all, or it will not cancel out the voice in the mix track, so that rules out using spectral processing tools.

A much easier option is to find a second copy of the song and simply edit the two songs together, missing out the parts that have the unwanted vocal. Or easier still, let the artist benefit a little and buy the CD.