Isolate Drums, Bass, and Piano Tracks with LALAL.AI

New innovative technology for high-quality stem separation of various musical instruments from songs and videos.

Isolate Drums, Bass, and Piano Tracks with LALAL.AI

We are introducing the most requested and highly anticipated feature - the possibility to isolate musical instruments in songs and video clips. In addition to splitting tracks into vocal and accompaniment stems, you can now use LALALAI to extract drums, bass, and piano from the mix.

In order to isolate instruments, go to the LALAL.AI site. Within the track upload section, click the Vocal and Instrumental button to open the Stem Separation Types list. Select the Drums, Bass or Piano option in the drop-down menu, then proceed to upload the audio or video file you want to extract the instruments from. The selection can always be changed after the track is processed.

Once the stem splitting is finished, you receive two tracks  - one containing only the instrument of your choice, the other with the rest of the music, background sounds, vocals (if there are any).

We’ve leveraged our superior neural networks and reached an unprecedented separation quality of music instruments isolation.

New neural networks for new splitting possibilities

In order to implement the musical instruments isolation, the LALALAI team used the innovative approach that gave us an unmatched separation quality to vocal and instrumental stems. Because of that, we were able to successfully avoid the quality reduction problems common among other music source separation solutions. One of such issues is the phase coordination of input mix and output stems - a lack thereof leads to extremely unpleasant sound artifacts.

The size of the data used to train new neural networks is measured in terabytes. Since individual instruments isolation is a more complex task than vocal and instrumental separation, it requires a more careful approach to processing, as well as significantly more data and compute resources.

For training, we’ve used some of the most robust servers capable of performing hundreds of trillions of operations per second. To make it easier to grasp the magnitude of power they provide, we can draw the following comparison - if our neural networks were trained on a top-end gaming PC instead of the servers, the training would take more than 7 years!

Try the new instruments isolation feature right now on the LALAL.AI site! It’s available without extra charge for all users, free and premium alike.