Learn
Everything we have learned about splitting songs into stems — how the models work, which mode to pick, how to fix the artifacts you will run into, and how to use the results in real projects. Written for musicians, producers, teachers and content creators.
Separation models do not “unmix” a song. They estimate how much of every frequency at every moment belongs to each instrument, then rebuild four or five new files from those estimates. Understanding that one idea explains almost every quirk you will hear in your stems.
Read the guideThe separation itself takes two minutes. Everything that makes a karaoke track actually work in a room full of people happens in the twenty minutes afterwards: killing vocal ghosts, restoring the low end, and leaving headroom for a live microphone.
Read the guideEvery additional stem is another chance for the model to guess wrong. Pick the fewest stems that answer your actual question, and your audio will sound better.
Read the guideSeparation artifacts are not random. Each one has a specific cause in how the model assigns energy, and each has a repair that works — plus several popular repairs that make things measurably worse.
Read the guideSeparated stems are sample-accurate copies of the same source, so they will line up perfectly if you import them correctly — and drift audibly if you do not.
Read the guideThe technology is neutral; the use is not. Private practice sits in comfortable territory, publishing an isolated vocal does not, and most real-world cases fall somewhere in between depending on which of two separate copyrights you touch.
Read the guideIsolating a part removes the hardest step in learning by ear: hearing it in the first place. Used well, that turns weeks of guesswork into an afternoon.
Read the guide