Acapella Extractor

Remixes and mashups. A vocal over a new beat is the oldest move in DJ culture, and for decades it depended on an official a cappella pressed on the B-side of a vinyl single, or a leak. Most songs never had one. Separation makes one for any track you own.

Learning a vocal part. A melody is far easier to hear without the band, and harmonies are easier still: with the instruments gone, the stacked voices in a chorus stop blurring into the guitars and you can pick out the line you are meant to sing.

Studying a production. Producers and engineers solo the vocal to hear what was done to it — how much compression, how bright, how much reverb, where the doubles and ad-libs sit. It is the closest you can get to opening the session file.

Transcribing lyrics. A word that is unintelligible under a wall of guitars is often perfectly clear on its own.

The old way to isolate a vocal was phase cancellation: take an instrumental version of the song, line it up to the sample, flip its polarity and add it to the original, so everything but the voice cancels. It works brilliantly when you have an instrumental from the same master, and not at all otherwise — a different mix, a different master or a lossy copy leaves a smeared mess.

This page needs only the song. A neural network, HT-Demucs, trained on multitrack studio sessions, estimates which parts of the sound are a voice and returns them as their own stereo track. How clean that track is depends on the recording more than on anything you do here:

• A dry, upfront vocal separates best — most modern pop, hip-hop and R&B.
• Heavy effects are harder: a vocoder, extreme distortion or a voice chopped into a synth can be read as an instrument.
• Instruments that sing — a lead guitar, a sax, a violin playing the melody — can leak into the vocal stem in places.
• The source file matters. Lossless or high-bitrate audio keeps the breath and consonant detail the model relies on; a low-bitrate rip loses it first.

An extracted vocal is a very good starting point, and a few minutes in any audio editor or DAW takes it the rest of the way:

Gate the gaps. Leftover instruments are most audible when nobody is singing. A noise gate, or simply cutting the silences by hand, removes most of what remains.

Cut the low end. Below roughly 80–100 Hz a voice contributes little, while kick and bass bleed live there. A high-pass filter cleans it up without changing how the vocal sounds.

Soften the very top. A gentle high-shelf cut tames cymbal wash, at the cost of a little air. Go lightly.

Mix it, do not just place it. Over a new track, the vocal will usually need its own reverb and some compression to sit naturally — a stem from someone else's mix was balanced for a different arrangement.