The Evolution of Source Separation
Audio Source Separation is the Holy Grail of audio engineering. For decades, mixing a song was like baking a cake—once the flour, eggs, and sugar (vocals, drums, bass) were mixed and baked, you couldn't separate them back out.
Phase Cancellation was the old trick. If you had the exact instrumental, you could invert its phase and play it over the original song to "cancel out" the music, leaving vocals. But you rarely have the exact instrumental.
AI & Deep Learning changed everything. By training neural networks on thousands of songs where the stems were known, the AI learned to recognize the specific spectrographic signature of a human voice versus a guitar or synthesizer.
Our tool uses Demucs, an open-source architecture from Facebook Research. It treats the audio signal not just as sound, but as a complex pattern, effectively "unbaking" the cake to give you back your ingredients.