Spleeter: A New Tool to Remove Vocals from a Song

Removing the vocals from a song is probably one of the first processes every music enthusiast wants to learn. Over time, tools designed for this have appeared, but they generally cost hundreds of dollars. The people at the streaming service Deezer decided to launch an alternative in the form of Spleeter, a source separation engine that, in addition to removing vocals, can also isolate other specific elements. The tool is very new; however, some results published on the Web are impressive.

You like a song. You obtain a digital copy. You open it inside an audio editor. A change here, an adjustment there… and then you take another song. You start cutting, you modify pitch, amplitude, and time. You decide to merge different parts, you listen and go back to the editor. Back and forth until a mashup is born. You upload it to YouTube (which will surely start screaming about copyright), and if you are a bit lucky, it will go viral.

Sounds good, right? The thing is this is an ideal case where everything works perfectly. In real life, the user must struggle to get good samples, and sometimes that requires something as straightforward as removing the vocals from a song. In other words, separating sources. Some enthusiasts have gotten good results with Audacity (emphasis on 'some', because many comments describe an 'underwater effect' after processing), and there are plugins capable of isolating voices, but they are not free (e.g., iZotope RX 7 costs $399 in its standard edition). Today, the people at Deezer offer us a third option: Spleeter.

How to remove the vocals from a song with Spleeter

What makes Spleeter different? Machine learning. Deezer trained the algorithm using its catalog, a resource that other researchers and projects simply cannot access due to copyright issues (not to mention time and energy), and by sharing its work with the community, they remove that feeling of 'unfair competition'. The two main details that currently limit Spleeter's reach are a relatively complex installation process, and the lack of an interface. In the specific case of Windows, you first need to install the full edition of the Anaconda platform, and then a git client. Spleeter is available in two versions, one for processors, and another that takes advantage of graphics cards (Nvidia hardware needs a compatible CUDA driver).

The three main Spleeter commands in the console are separate to separate sources, train to train an alternative separation model with your own dataset, and evaluation, which tests the model against the MusDB set. Under the right conditions, Spleeter has proven to be very fast: On a 32-core Xeon server with a GTX 1080, Spleeter chewed through the three and a half hours of musDB in 90 seconds. If the plan is to process a couple of tracks at a time, I figure you won't have problems with your hardware.

The first Spleeter-based mashups have already appeared on the Web. Some are very interesting, while others should remain in the dark. Andy Baio from the Waxy portal gives us a fairly precise idea:

The rest is a matter of waiting for a more user-friendly environment for the average user. Spleeter's potential to remove vocals from a song and separate other sources is gigantic, especially being free and open source.

Official announcement: Click here

Source: Waxy.org