Neural vocal remover for your local music stash

What it does

vocal-remover is a PyTorch-based utility that takes a standard audio file and splits it into two files: one containing the vocals and one containing the instruments. It runs entirely offline on either a CPU or GPU. The project also includes a training pipeline, so you can retrain the model on your own dataset of mixed songs and their isolated instrumental counterparts if the pre-trained weights don’t suit your genre.

The interesting bit

The README is admirably terse: it cites a half-dozen influential source-separation papers—from U-Net to DenseNets to complex masks—but never actually says which architecture the code implements. You get test-time augmentation and an experimental post-processing mask that tries to clean up the instrumental track based on residual vocal energy, but the underlying model remains a bit of a black box.

Key highlights

• Runs fully offline on CPU or GPU; no cloud service or API key required.
• Splits a single input file into separate instrumental and vocal WAV outputs.
• Supports test-time augmentation (--tta) to squeeze out cleaner separation.
• Includes an experimental post-processing option that masks instrumental regions based on detected vocal volume.
• Offers a training script for custom datasets of mixed audio and isolated instrumentals.

Caveats

• The --postprocess flag is explicitly marked experimental; the README warns it can cause problems and should be disabled if anything sounds off.
• The README cites several research papers but does not clarify which specific architecture or model the tool actually uses, so evaluating the approach requires reading the code.

Verdict

Worth a look if you need quick, local vocal separation for remixing or karaoke and don’t mind a little mystery under the hood. Skip it if you need production-grade transparency, detailed documentation, or guaranteed state-of-the-art benchmarks.