How to Remove Vocals From a Song (Without Uploading It)
Removing a vocal is a short job once you know which file to feed the model and how to check the result. This walkthrough covers both, plus the cases where a browser-based tool is the wrong choice.
6 min read · Updated 2026-11-21
What you need
A desktop browser with WebGPU — current Chrome, Edge, or Safari 17 and later. Without it the model has no GPU to run on, and a CPU fallback would take far longer than the song itself.
The best copy of the song you have. A 320 kbps MP3 or a WAV is ideal. A low-bitrate rip gives the model less to work with, and cymbals smear into the range where the voice lives.
No account, no email. The tool starts as soon as its 67 MB model is cached by the browser.
Step 1 — Start with the best source file you have
Separation is a subtraction, so everything already in the source stays in the output. Clipping, a mono fold-down, or heavy MP3 artifacts above 15 kHz all survive the process.
Stick to 44.1 kHz. Upsampling a 22 kHz file does not create information the model can use, it only makes the file bigger.
Video files are fine. MP4, MOV, MKV, AVI, WMV, FLV, and TS are decoded first, and only the audio track is sent to the model — the picture is never touched.
Step 2 — Run the separation on your device
Drag the file onto the page or click to browse. The first run downloads the model once; after that it loads from the browser cache and starts in seconds.
Expect roughly the length of the track. A four-minute song takes one to two minutes on a recent laptop GPU, longer on integrated graphics. Tracks past 20 minutes are allowed, but the page warns you before the run starts.
Keep the tab open while it runs. The computation happens in that page, not on a server you can walk away from.
Step 3 — Listen for residue before you export
Play the instrumental on headphones. Backing vocals, ad-libs, and reverb tails are what usually survive; a dense chorus hides them better than a quiet verse, so check the quiet part first.
Play the vocal stem too. If it sounds thin, the model had to remove everything overlapping the lead, including parts you wanted. That is expected behaviour, not a failed run.
Export both stems as 192 kbps MP3 at 44.1 kHz. That is enough for rehearsal, karaoke, and most remix work; this tool does not output lossless WAV.
Why local processing matters
Most online removers upload your file, process it on a server, and hand back a download link. A copy exists somewhere for as long as their retention policy allows.
The free mode here has no upload endpoint at all. The file is decoded inside the tab and the model runs on your GPU, so there is no copy to retain and no account to tie it to.
The same design has a cost: you cannot close the laptop and come back to a finished job. The compute is happening on the machine in front of you.
Common problems
Most failed runs fall into one of these five buckets. Check the bucket before assuming the model is broken.
- “Local separation isn't available on this device” — WebGPU is missing or disabled. Update the browser, check chrome://gpu, or use cloud separation instead.
- Some vocals remain in the instrumental — backing vocals and harmonies overlap the lead. Try a cleaner source before blaming the model.
- The vocal stem sounds thin — everything overlapping the voice was removed with it, which is the same operation that made the instrumental clean.
- The run feels stuck — long tracks on integrated graphics can take several minutes. The progress bar moves per chunk, not per second.
- No MP4 appeared for a video — the browser could not encode AAC. The two MP3s are still produced, and the result panel says which step stopped.