A lightweight alternative to AI vocal separation
Demucs and Spleeter are dedicated source-separation systems that use trained models to estimate vocals and accompaniment. Running those models reliably requires a different processing architecture, larger downloads or server compute, and a clear audio-upload and retention policy.
This first version takes a lower-cost path: it subtracts the shared center signal from a stereo file. It can reduce centered lead vocals in some mixes, but it can also remove centered bass, drums, synths, or other instruments. The output is not an isolated instrumental stem.
When center reduction works best
Use a stereo file where the vocal is mixed near the center and the accompaniment has useful left/right differences. Older karaoke mixes and some studio arrangements may respond noticeably; heavily reverberated, doubled, panned, or mono recordings usually will not.
Processing happens locally with the browser Web Audio API. The selected file is decoded in memory, converted to a WAV download, and is never uploaded.
What the browser is actually calculating
A stereo recording stores a left channel and a right channel. This experiment creates a difference signal by reducing the audio that is shared by both channels. In a simplified form, the result behaves like left minus right. Sounds placed in the middle of the stereo field may therefore become quieter, while sounds with useful side information remain more audible.
That explanation also describes the limitation. A mixer can place many elements in the center, not only the lead vocal. Kick drum, bass, snare, piano, speech, and effects may all lose energy together. The method does not know which sound is “the vocal”; it only responds to the spatial relationship between channels.
A practical workflow for a better experiment
Start with a copy of the original file so the source remains unchanged. Choose a stereo track with a clear center-panned vocal and listen for left/right differences before processing. After creating the reduction, compare a short section with the original at a similar listening level. Check the beginning, a chorus, and a quieter section because a result that sounds promising in one passage may reveal artifacts elsewhere.
Use headphones or speakers that reproduce stereo separation, but remember that playback hardware changes what you hear. If the source is mono, both channels contain the same signal and the subtraction can remove almost everything. If the mix has wide vocal doubles or long reverb, the voice may remain as an echo or lose only part of its energy. Treat the output as an experiment, not a guaranteed karaoke track.
Privacy, copyright, and local files
The selected file is decoded in the current browser tab. No upload endpoint is involved, and the output WAV is assembled locally for download. Closing the tab removes the in-memory working data unless you separately save the downloaded file. As with every browser tool, other extensions or software on your device may have their own access; keep your browser and operating system up to date.
Only process audio you own or have permission to edit. A local workflow does not change copyright ownership or grant permission to redistribute a song. If you share an output, check the license and the platform rules that apply to the original recording.