For the 70s/80s/90s mixes I make, I’m always looking for original acapellas, or even better, the original multitracks of the recordings. And if you manage to find it, its price is in line with its rarity, up to hundreds of euros each.
But what about artificial intelligence? What if you train a computer model with the original recordings you have and let the model learn the sound of the different instruments and vocals? Can such a model filter the vocals from an existing song?
I put it to the test and made the mashup below without the original separate tracks of the songs. AI computed for a few hours and finally gave me an instrumental version of one song and the acapella of the other song. Fortunately, there is still a lot of creative manual work involved to arrive at the result below.