L−R vs Demucs: why AI vocal removal beats center cancel

The comparison starts from a simple reality: most pop mixes place the lead vocal equally in both channels, so subtracting one channel from the other removes it. That is the entire logic of L−R, and it is shown in a handful of Python lines. But the same logic removes any other centered element — bass, kick and snare are often placed there too — and the subtraction leaves a single mono channel, so stereo width is lost. Reverb on the vocal spreads across the stereo field and survives, leaving a ghost of the voice.

Demucs takes a different route. Its htdemucs model, built on a Hybrid Transformer architecture and released under the MIT license, learns what vocals, drums, bass and other instruments sound like from many songs, then estimates each source from the mixed audio. Because it separates by sound character rather than channel difference, centered instruments stay in the backing, and vocals can be pulled out as their own track. The official repository was archived on January 1, 2025, with only important fixes continuing in the author's fork.

Practical limits remain on both sides. L−R is nearly weightless but produces thin, mono results and cannot handle mono recordings at all. Demucs is heavier — about 7GB of GPU memory by default, though --segment or CPU mode can reduce that at the cost of speed — and it can still leave vocal traces on deep echo, live crowd recordings or old mono sources. Choirs and harmonies land in the vocals stem, so lead-only extraction is not its strength. For readers who want to skip setup, a browser-based service, Voice Remover, runs htdemucs in the backend and returns MP3 backing and vocal tracks.