In this video I compare three pieces of software that separate vocals from instrumental tracks: Spleeter, DEMUCS and LALAL.AI. Besides the quality of the separation there are a few relevant considerations. Spleeter and DEMUCS are open-source and free, while with lalal.ai you can process 10 minutes of material after which you have to pay. DEMUCS and Spleeter can both be run locally or in Colab, while you have to use the website of lalal.ai to perform audio separation. It also just splits the songs in two tracks: vocals and instrumentals, while Spleeter and DEMUCS can split them into more tracks (drum, bass, vocals and rest). Lastly Spleeter can be trained on a new dataset. For making backing tracks (removing the guitar track) DEMUCS is by far the best option, as it performs vocal isolation much better with heavily distorted guitars than Spleeter, and the drums sound much more clear. The disadvantage is that the vocal track can be a bit crackly, so then it can be supplemented or substituted with the vocal track from Spleeter.

Song used: [ Ссылка ]

lalal.ai: [ Ссылка ]
Spleeter: [ Ссылка ]
Spleeter colab: [ Ссылка ]

Demucs: [ Ссылка ]
Demucs colab: [ Ссылка ]

EDIT:
Demucs released a new version. A colab of it can be found here: [ Ссылка ]

It sounds better than the version shown in the video, especially the bleed between the vocals and the others track is much improved!

свернуть