LALAL.AI has rolled out an update to its Perseus neural network, adding Acoustic Guitar, Electric Guitar, and Piano stem separation, which offers musicians and producers a new level of precision when isolating these instruments from full mixes.
Previously, these stems were extracted using the Orion network. With Perseus now handling the task, users can expect more refined results, thanks to the model’s advanced architecture: Perseus operates on transformer-based models, a method also used in AI language processing, like OpenAI's ChatGPT.
This approach enhances the network’s ability to recognise and separate complex harmonic structures and makes it more effective at distinguishing individual instruments. LALAL.AI first applied this technology to vocal and instrumental separation and has now expanded its use to guitar and piano stems. According to LALAL.AI, "Perseus outperforms its predecessor by a remarkable one decibel across various metrics, translating to about 15% improvement in vocal extraction quality."
How to Try It
The new feature is accessible directly on the LALAL.AI website. Users can upload an audio or video file, select the desired stem from a drop-down menu (all Perseus-supported stems are marked with green dots), and let Perseus process the track. There’s also an option to switch between neural networks in the settings menu for comparison.
For those looking to extract cleaner instrumentals or dive deeper into song structures, these new stems offer another tool in the creative toolkit.
