Full-quality stems. No account, no per-track fee.Same model and same 192 kbps (or lossless WAV) as the paid tier — tracks up to 15 minutes.
Karaoke night, a cover, practice, a backing track for your set: upload the song and StemGrab hands you the instrumental as MP3, plus every separate stem. Free, no account, no watermark.
Your upload is deleted the moment the split is finished. Want WAV, play-along mixes or longer tracks? The full tool is on the home page.
Make an instrumental →45 instrumentals have been made through this exact page in the 40 days we've measured. Sitewide, across every upload in that stretch, the average song runs 4.0 minutes, the longest was 44.1 minutes, and most cluster between 2 and 4 minutes (35 of them) or right around the 1-minute mark (20) — short clips and full songs both show up. 97% come in comfortably under our limits, and not one upload in the batch landed within 10% of the ceiling before going over. The one that does bite: a 15-minute cap on any single upload. "Track is longer than 15 minutes — please trim" is the message people hit when they try a full album side or a DJ set instead of a song — 22 times in 40 days, one of the more common ways an upload gets turned away here. Chasing a backing track for a long set rather than one song? Split it into parts first; this tool is built for a single track, not a mix.
Pre-made karaoke tracks are re-recordings: different key, different feel, often behind a paywall. StemGrab strips the vocals from the original recording, so the groove, the sound and the energy stay exactly as you know them. And because the tool separates the whole mix, you also get the drums, bass, guitar, piano and other instruments as bonus tracks for practice or remixing.
For years the only free way to kill vocals was the phase-cancellation trick: take a stereo song, invert one channel and sum it to mono. Because the lead vocal usually sits dead-centre, it partly cancels out — but so does the bass, the kick and anything else panned to the middle, and the result is a thin, hollow, mono mess. It only ever half-worked.
StemGrab doesn't do that. Our separation model is trained specifically on voices, so it lifts the vocal out and leaves the drums, bass and instruments full and in stereo. You get a real instrumental that still sounds like the record, not a phase-mangled ghost of it. The same upload also hands you the isolated vocal, so nothing is thrown away.
For a karaoke or sing-along track the lead vocal is what you want gone, and on most modern, well-mixed songs it comes out cleanly. Two honest caveats so you know what to expect: backing vocals and harmonies woven through the mix can partly survive, since the model treats them as part of the "vocal" it's isolating and they're harder to place. And on dense, live or heavily reverbed recordings you'll hear a little more residue than on a clean studio mix — that's true of every separation tool, not just this one. If a faint vocal ghost bothers you, drop the isolated vocal stem back in at a low level: it fills the gap and masks any artefacts far better than a dead-silent instrumental.
Just want the isolated vocal instead of the backing track? Head to extract vocals from a song. Once you've got your instrumental, you can also practice with the backing track: mute your own instrument and play along with the rest of the band. An instrumental is audio; sheet-ready notes are something else. MP3 to MIDI converts the same song into a MIDI file with bass, guitar and piano each on their own track.
This is worth understanding, because it explains every result you'll get. The instrumental is not made by cancelling the voice out of the mix — it's rebuilt from the parts that aren't the voice. The model separates the song into six pieces (vocals, drums, bass, guitar, piano and everything else), throws the vocal piece away, and adds the other five back together. What you download is a mix of five separated stems.
Two consequences follow directly from that:
The upside of the same design: because the backing is reassembled from separated stems rather than filtered, the drums stay punchy and the bass stays whole. Nothing is notched out of the middle of the frequency range, so the instrumental doesn't get that hollow, phased sound that older tricks produce.
Some material is simply easier than other material. From easiest to hardest:
Two practical habits make more difference than people expect. First, start from the best copy you have: a purchased download or CD rip has detail that a low-bitrate file threw away years ago, and the model can only work with what's in the file. Second, check the loudest chorus, not the intro. Intros are sparse and always sound perfect; the last chorus with every layer playing is where you find out whether this song separates well. If it doesn't, you'll know in fifteen seconds instead of after a full render.
The output is a 192 kbps MP3 — fine for singing along, rehearsing or a backing track through a PA. Tracks up to 15 minutes are accepted, and your files are deleted automatically within 45 minutes of the job finishing.
MP3, WAV, FLAC, M4A, AAC, OGG, Opus, AIFF and WMA, and video files like MP4, MOV and MKV — we pull the audio out ourselves. Up to 2 GB and 15 minutes per track.
Usually well under a minute per song: the separation runs on a real GPU, not a queue of credits.
Yes, every split produces the instrumental and the isolated acapella, plus drums, bass, guitar, piano and other.
Harmonies and backing vocals baked deep into the mix are treated as part of the vocal the model isolates, so a little can remain on busy songs. Nudge the isolated vocal stem back in at a low level to mask it, or try a cleaner studio version of the track.
Deleted straight after processing; results wiped within 45 minutes. See the privacy policy.