Three different jobs people call "removing vocals"
The phrase covers at least three tasks that need different tools:
- Make a karaoke track — remove the lead vocal and keep the instrumental.
- Isolate the voice — keep the vocal and discard the music, for sampling or analysis.
- Separate stems — pull apart vocals, drums, bass and other instruments into separate files.
What is possible depends entirely on which of these you need and how the original was produced.
The old trick: phase cancellation
Before AI, vocal removal relied on a quirk of stereo mixing. On many older recordings the lead vocal is placed dead-centre, meaning it appears equally in the left and right channels. If you invert one channel and sum it with the other, the centre cancels out — and the vocal disappears with it.
This is the technique behind classic "vocal remover" tools. It works surprisingly well on some 1960s and 1970s stereo mixes, and it is instant. But it has hard limits:
- Anything else in the centre — kick drum, bass, some guitars — vanishes too, so the backing track sounds thin and hollow.
- Modern productions often spread the vocal with effects, so it is not purely centred and does not fully cancel.
- Mono recordings cannot be processed this way at all.
The modern approach: AI stem splitting
AI stem splitters take a different route. Instead of a single phase trick, a neural network is trained on huge amounts of music to recognise what a voice sounds like versus what instruments sound like, then it estimates each source separately. The result is far cleaner than phase cancellation, and it works on modern music and on mono files.
It is still an estimate, not a perfect extraction. Where a voice and an instrument occupy the same frequencies, the model has to guess, and you will hear artefacts: a little bleed, a slightly smeared cymbal, a faint ghost of the vocal in the backing track. On a sparse mix — a singer over a simple guitar — the result can be genuinely impressive. On a dense, wall-of-sound production, it will be rougher.
How the Offex Audio tool fits in
The Voice / BGM Isolation tool is built for the first two jobs: keeping the voice, or keeping the music. It is a server-side AI feature, which matters for two reasons. First, the model is far too heavy to run comfortably in a phone browser. Second, because it runs on a server, the tool depends on that processor being online.
That is exactly what the status badge on the home page is for. When a processor is connected the badge shows a green Live; when it is not, the badge reads Not connected and the separation tool explains itself instead of failing silently. The other two tools — Remove Silence and Enhance Audio — are unaffected, because they run on our secure server too, with a lightweight in-browser fallback when the server is unreachable.
Practical tips before you separate
- Start from the best source you have. A lossless WAV or a high-bitrate file separates far more cleanly than a low-quality MP3.
- Expect to clean the result. Run the isolated voice through the enhancer afterwards to reduce bleed and level it.
- Respect the rights. Removing vocals from a song you do not own does not make the result yours to publish. Use these tools on your own recordings or material you are licensed to use.
Frequently asked questions
Can I remove vocals from a song online for free?
You can try. Simple vocal removal works on some older stereo recordings, but modern, densely produced songs need an AI stem splitter. The Voice / BGM tool on the home page shows whether a server processor is connected before you rely on it.
Is removing vocals the same as making a karaoke track?
Usually, yes — a karaoke track is a song with the lead vocal removed. The catch is that the backing track left behind is never exactly the original instrumental, because the split is an estimate.
Why does vocal removal sometimes sound hollow or thin?
Because voices and instruments share frequencies. Removing the vocal removes some of the music with it, and what is left can sound thin. AI splitters reduce this, but they cannot eliminate it entirely.