vocalremover.org vs Extractune, tested on one song

Our test: one song, scored against its own stems

Listening tests are hard to repeat, so we wanted a number. That takes a song whose separate tracks are public: "Just Lucky, I Guess" by Mr_Yesterday, 3:18 long and released under CC BY 4.0. Its artist also published the original vocal track and the instrument tracks it was mixed from: bass, drums and a baritone sax. The vocal track, with its effects, is what a perfect vocal remover would give back as the acapella, and the three instruments added together are the perfect instrumental.

On 6 November 2026 we split the whole song with both tools on one computer (Windows, Chrome, 12 processor cores, 16 GB of memory, WebGPU). Extractune used its free split: HTDemucs, running on the graphics chip inside the browser. vocalremover.org was used the normal way, without an account, which uploads the song to its server. We downloaded each tool's vocals and instrumental as WAV files and scored them.

Song: “Just Lucky, I Guess” by Mr_Yesterday, licensed CC BY 4.0.

What the scores mean

SI-SDR (scale-invariant signal-to-distortion ratio) is the usual score in source-separation research. It splits a tool's output into the part that matches the real track and everything else: the instruments left in an acapella, the voice left in an instrumental, and any warble the model adds. The score is the ratio of the two in decibels, so higher is cleaner, and 3 dB more means everything else carries half the energy. Plain SDR is the same idea without the scaling step, so it also counts a part that is only quieter or louder than the original as error.

The bottom row is the starting line: the unsplit song scored as if it were the vocals (0.64 dB) or the instrumental (−2.06 dB). Both tools moved a long way from it. On this song Extractune's vocals scored about 1.0 dB higher (10.76 dB against 9.79 dB) and its instrumental about 1.9 dB higher (8.31 dB against 6.38 dB). By plain SDR the vocals are close to level (8.60 dB against 8.42 dB), and the instrumental gap stays at about 1.5 dB. The gap between the two tools is far smaller than the step either one makes from the unsplit song.

This is one song on one computer. A song with a louder mix, more reverb or several singers can come out differently, and vocalremover.org can change its model at any time. Treat it as one honest data point rather than a ranking, and re-run it on your own songs: the method is below.

How we scored it

Every file was decoded by the same decoder, Chrome's own, and compared with the artist's stems: the vocals against the vocal track with effects, and the instrumental against the sum of the bass, drums and sax tracks.

Each channel is lined up with its stem first, by cross-correlation within 2,000 samples either way, because a split part can start a few milliseconds early or late. The first and last 2,048 samples are left out, and the left and right channels are scored on their own and averaged. Both SI-SDR and plain SDR are reported.

The lining up mattered here: Extractune's downloads of this MP3 start 1,015 samples (23 ms) later than the song, while vocalremover.org's line up exactly. You won't hear 23 ms when you practise or sing along, but a producer lining the parts up against the original in a DAW would, so we're looking into it. It doesn't change the scores, because each part is lined up before it's scored.

Anyone can repeat it: split a song whose stems are published (ccMixter has many under Creative Commons) with both tools, download the parts as WAV, and score each against its stem the way described above. Try it on your own phone or computer, since speed depends on the device.

Speed: about the same

Extractune split the 198-second song in about 27 seconds on this computer, 7.39 times faster than the song plays, with nothing uploaded. vocalremover.org took about 24 seconds from choosing the file to the result showing, upload included; its own site says processing usually takes about 10 seconds.

On a computer whose graphics chip the browser can use, that's a draw. Elsewhere it isn't: Extractune runs on your own device, so a phone takes longer (the best free vocal removers guide has split times measured on visitors' phones and computers), and the very first split downloads the 90MB engine once. vocalremover.org's time depends on your upload speed and its servers instead, so on an older phone with a good connection it's likely to be the quicker one.

What leaves your device

With vocalremover.org, the song is uploaded to its server, split there, and the two parts come back to your browser. That's how its vocal remover works, and it's why there's nothing to download first.

With Extractune, the song never leaves your device: extracting, splitting, the mixer and every download run in your browser. The only feature that sends audio anywhere is Get the words (a transcript and song ID), and only when you tap it. For an unreleased demo, a client's track or a recording with people in it, that difference can matter more than a decibel.

What you get from each

vocalremover.org's vocal remover gives two tracks: a karaoke version with no vocals, and the vocals on their own. It has a separate Splitter for vocals, bass, drums and the rest (its site says that one usually takes about a minute), and a shelf of small tools around them: a pitch changer, a key and BPM finder, a cutter, a joiner and a voice recorder.

Extractune gives four parts from one split: vocals, drums, bass and everything else, on one timeline with a fader, mute, solo and download for each, and one-tap mixes for an acapella, an instrumental, a drumless or bassless track, or the drums alone. Downloads come as MP3, M4A, WAV, FLAC or WebM, one part at a time or all four in a zip.

It also starts from a video (MP4, MOV, WebM or MKV, up to 500MB): split its sound, then save the video back with only the voice, only the music, or any mix of the parts. And the parts go straight into practice: slow down to 50% without changing the pitch, change the key by up to 12 semitones, and loop a section.

Side by side

vocalremover.org details are from its public website and app listings, checked in November 2026. Plans and limits change, so check vocalremover.org's own site for the current ones. Extractune isn't affiliated with vocalremover.org.

Where each one wins

Which one to pick

Pick vocalremover.org for a quick vocals-and-instrumental split on a device that's slow to split on its own, or for its other small tools, when uploading the file is fine.

Pick Extractune if you'd rather the song stayed on your device, you want all four parts or the video back, or you want to practise with what you split.

Is vocalremover.org safe?

We can't vouch for another site, so here's what we could check, on 6 November 2026. Its vocal remover needs no account. The song you pick is uploaded to its server to be split, which is how the tool works rather than a flaw. Its privacy policy (vocalremover.org/privacy-policy) covers visitor log files, cookies and advertising partners; we found nothing in it about how long an uploaded song is kept or when it's deleted.

For a song that's already public, that may not matter. For one you'd rather not hand over, like an unreleased demo or a private recording, pick a tool that splits on your device.

Questions