Demucs is an open neural network from Meta's research lab for splitting music into separate tracks. It's considered one of the most accurate models available: vocals come out without a muffled, underwater quality, and the instrumental is left with almost no trace of the voice.
Two modes are available. "Vocals and Instrumental" produces two tracks — a clean voice and the backing accompaniment, perfect for karaoke, covers, or remixes. "All Instruments" breaks a track into six parts: vocals, drums, bass, guitar, keys, and everything else.
You can upload audio or video up to 10 minutes long. A 3-minute song costs 16 credits.