Andromeda is LALAL.AI's latest model for splitting audio stems like vocals or instruments. It’s built on the same transformer-style design as the earlier version but trained on way more data, about four times as much. They say it’s about 40% faster and pulls cleaner stems with around 10% better signal-to-distortion ratio.
It’s already built into tools like Stem Splitter, Vocal Splitter, Voice Cleaner, and Echo & Reverb Remover. It’s also heading to mobile and plugin formats.
They’ve been working on this model for over nine months. It builds on six earlier models like Rocknet, Phoenix, and Perseus. Like before, it doesn’t just look at audio as a flat stream, it breaks it down by time, tone, and frequency to handle all kinds of tracks, even ones with soft or loud parts.
One of the big changes is that it no longer makes you pick between getting more detail or less stem bleed. Now it just aims to do both.
You get the audio back as separated vocal, instrument, or noise tracks in stereo up to about 22 kHz.
If you'd like to access this model, you can explore the following possibilities:
