Free AI Vocal Remover

Most songs you love exist only as a single finished file. The studio kept the individual tracks — the lead vocal, the drum bus, the bass line, the guitar — locked away, and all you get is one stereo mix. This page splits that mix back apart. Drop in a song and OwnSound hands you two clean files: an instrumental with the voice taken out, and the isolated vocal on its own. There's no account, no watermark, and no cap on how many songs you run through it.

The reason it can stay free is that the AI — Meta's Demucs model — runs on your own computer instead of a rented server. That's a genuine trade-off. A cloud vocal remover finishes a track in seconds, while this one needs minutes and works best on a desktop with a decent graphics chip. In exchange you keep your music to yourself, and there's no subscription or daily limit waiting to cut you off halfway through a project. For a band rehearsal track or a remix idea at 2 a.m., that's usually the better side of the bargain.

Remove Vocals From a Song in Three Steps

1. Pick your file. MP3, WAV, FLAC, OGG, M4A — if your browser can play it, the tool can usually process it. The first time you use it, the AI model downloads once (about 170 MB) and then stays cached, so later visits skip straight to separating.
2. Let it separate. The model works through the track in short windows, scoring each one as drums, bass, vocals, or "other" — guitars, keys, and everything else. A progress bar ticks up as each chunk finishes, and you can cancel at any point.
3. Grab both stems. Two players appear when it's done: the instrumental and the isolated vocal. Preview each one in the page, then download the ones you want as WAV files.

Make a Karaoke Track From Any Song

Karaoke versions of older or less famous songs often simply don't exist, and the ones that do get pulled from streaming services on copyright claims all the time. Stripping the vocal yourself sidesteps both problems. The instrumental you get back is the full arrangement — drums, bass, guitars, keys — with just the lead singer removed. It won't be a perfect match for a studio instrumental, because backing vocals that were layered into the vocal stem usually come along for the ride, but for a night of singing in the living room it's more than good enough.

There's a small art to choosing a track that will separate cleanly. A song with a single lead vocal sitting up the middle, and a clearly separated band behind it, comes out nearly spotless. A dense wall of harmonized vocals is harder, because the model has to decide where "voice" ends and "texture" begins. If a studio version separates messily, try a live acoustic take — a singer with just a guitar often splits far better than the full album cut with all its overdubs and effects.

Isolate Vocals and Grab a Clean Acapella

The second output is the acapella: just the voice, lifted out of the full mix. Remix producers use this as raw material — chop a line, pitch it up a fifth, layer it over a new drum pattern. If you've ever hunted for a clean acapella of a song that was never officially released as one, you know how scarce they are, and how fast the sites that host them get taken down. Making your own from the file you already have in your library is the dependable way to get one that won't disappear.

The vocal output isn't a studio-grade isolated track, and it's worth setting that expectation up front. Because Demucs rebuilds the voice rather than just muting everything else, you'll sometimes hear a ghost of the snare or the reverb tail bleeding through, especially on loud, heavily compressed mixes. For a hook dropped over a new beat, that's almost always inaudible once it's sitting in context. For a dry lead you want to feature on its own, a little EQ or a noise gate will clean it up. Either way, it beats not having the vocal at all.

How a Modern Vocal Remover Works

The old way to remove vocals was to subtract the left and right channels. A centered lead vocal is identical in both, so flipping one side's phase cancels it — along with the kick, the snare, the bass, and anything else panned dead center. What's left is a hollow, phasey shell of a song. It's the sound of every budget karaoke CD from the nineties, and it's exactly why the trick has been abandoned by anyone serious.

Demucs approaches the problem from a completely different angle. It's a neural network trained by Meta on a huge catalog of music where the separate stems were available, so it learned to recognize a voice by what it sounds like — the vibrato, the sibilance, the breath — rather than by where it sits in the stereo field. When you feed it a song, it listens in short windows and scores each one as drums, bass, vocals, or "other," then rebuilds each part independently. That's why it can pull a vocal out of a track that's entirely mono, something the phase trick could never touch.

Get a Cleaner Instrumental From Better Source Files

• Start with the best file you own. A 320 kbps MP3 separates noticeably better than a 128 kbps rip, and a lossless FLAC or WAV is better still. Compression artifacts read as noise to the model, and noise ends up smeared across every stem.
• Prefer the official mix over a live bootleg. Studio tracks have clean, separated elements. A phone recording of a concert smears everything into one blur, and no model can untangle that.
• Keep songs under about seven minutes. Long tracks use more memory, and on a phone they can simply run out and fail partway through.
• Use Chrome or Edge on a desktop with a GPU. WebGPU support makes a real difference to speed; Firefox works but is slower, and Safari trails behind.
• Listen before you download. Preview both stems in the page, then keep only the one you'll actually use — the files are lossless WAVs and a whole album adds up fast.

How Long Removing Vocals Takes on Your Computer

Because the model runs on your hardware, the time it takes depends entirely on what you're using. On a desktop with a graphics card, expect roughly two minutes of processing for every minute of audio — a three-and-a-half-minute pop song lands in about seven minutes. On a computer with no GPU, the same song can stretch to twenty minutes or more, because the model falls back to plain CPU math and grinds through one chunk at a time. A progress bar tracks each chunk as it goes, so you can see roughly where it is in the song and cancel early if it's dragging.

The very first visit adds one extra step: the roughly 170 MB model has to download once before anything else happens. That's a one-time cost, and after it's cached the page loads straight into separating. If you're on a laptop running on battery, or on a phone, plug in and expect it to take longer still. The model also has to hold the entire track plus its own weights in memory at the same time, which is why a desktop is the comfortable place to run it, and why closing other tabs helps.

Inside the Separation: Four Stems, Two Outputs

Demucs actually splits a song four ways — drums, bass, vocals, and a catch-all "other" bucket that holds guitars, keys, synths and everything else. OwnSound then merges three of those into the instrumental, so you end up with exactly two files: the full backing and the lead vocal. We keep it at two on purpose, because that's what karaoke singers and remixers actually ask for, and it keeps the page from turning into a mixing console. If you want the drums or bass on their own, that's a job for a full stem splitter, which hands you all four instead of recombining them.

The split isn't always clean at the edges, and that's normal for every tool in this category. Reverb and delay tails often ride along with the vocal, since the model hears them as part of the voice, so a ballad's echo can linger faintly in the acapella. A singer who doubles a guitar line can pull a bit of guitar into the vocal stem too. LALAL.AI, Moises and PhonicMind all produce the same kinds of artifacts, because they all work from the same underlying idea. In a finished mix, they're almost always inaudible.

What People Use a Vocal Remover For

OwnSound vs LALAL.AI, Moises, PhonicMind & Server-Based Tools

If you've shopped around, you've likely come across LALAL.AI, Moises and PhonicMind. They're all legitimate, and they all run the same core idea — a neural network separating audio into stems — but they host the model on their own servers. That's why they're fast, and it's why they charge or cap you. The table below lays out the honest differences. The short version: OwnSound trades raw speed for being free, unlimited, and keeping your files on your side of the screen.

	OwnSound	LALAL.AI	Moises	PhonicMind	Server tools
Price	Free, no limits	Paid packs or subscription	Freemium (~5 tracks/mo)	Paid per track	Freemium or paid
Account required	No	Yes	Yes	Yes	Usually yes
Your audio uploaded	No	Yes	Yes	Yes	Yes
Where it runs	Your device	Their servers	Their servers	Their servers	Their servers
Speed	Minutes (local AI)	Seconds	Seconds	Seconds	Seconds
Separation model	Demucs (local)	Proprietary	Proprietary	Proprietary	Varies
Stems available	2 (instrumental + vocals)	Up to 10 stems	Up to 5 stems	2–5 stems	2–10 stems
Output formats	WAV	MP3, WAV, FLAC	MP3, WAV	MP3, WAV	Varies
Works offline	Yes (after first download)	No	No	No	No
Practical length limit	~7 min	None	5 min (free)	~10 min	Varies
Watermark on output	No	No	No	No	Sometimes
Best on mobile	Desktop recommended	Yes (app)	Yes (app)	Yes (web)	Yes