Vocal remover online

Remove vocals, pull karaoke beats and isolate stems with AI

Updated: Sep 1, 2026

Paste a video or track link

Paste a link to separate it directly, no need to download the file first.

Drag & drop a file here, or click to choose

mp3, wav, m4a, mp4… up to 200MB

No file handy?

Free with a limit; past that it costs Energy. The on-device engine never sends your file anywhere, the server engine does upload it to process.

Still hearing vocals in the instrumental?

Noisy live takes, stacked harmonies, dense arrangements and heavily compressed files are the cases an automatic model cannot finish. Our team separates those by hand.

Send us the track

Music separation splits an already-mixed recording back into separate layers: vocals in one file, the instrumental in another, and if you want, drums, bass and the remaining instruments each on their own. An AI model listens to the whole track and estimates, at every moment and every frequency, which layer that energy belongs to. Upload a file or paste a link, choose how many layers, then download each one.

What is music separation?

Music separation splits an already-mixed recording into separate audio layers, each one called a stem. From a single MP3 you get the vocal as one file and the instrumental as another, or in more detail drums, bass, piano and the remaining instruments each on their own. This is not the multitrack session coming back: that session was mixed down to two stereo channels long ago and no longer exists inside the file you have. An AI model reconstructs each layer from what it learned across a very large number of other recordings, so the result is always a close approximation, never the original.

What is a karaoke backing track?

Pulling a karaoke backing track means removing the singer's vocal from the original recording while keeping everything else (instruments, drums, bass, strings) intact. With the 2-stem mode, the instrumental is separated out of the original recording itself, not rebuilt as MIDI and not re-arranged from scratch.

Should you pick 2, 4 or 5 stems?

2 stems (vocals and instrumental) is the right choice for karaoke, covers and vocal practice: fewer layers means each layer comes out cleaner. Pick 4 stems (vocals, drums, bass, other) when you want to hear how the track was built, or to lift the drum groove on its own. 5 stems pulls piano out of the "other" group, which helps on ballads and lounge material, but the more layers you ask for, the more audible the mistakes at the boundaries between them become. The on-device engine handles 2 and 4 stems; 5 stems needs the server engine (spleeter) on a paid run.

Why does one song separate cleanly and another not?

Separation is an estimation problem, not a recovery of the original session. The more open the mix and the more clearly the vocal sits apart from the instruments, the tighter the result. A dense arrangement, doubled or stacked harmonies, a guitar or synth sitting in the lead vocal's frequency range, or a source file already crushed to a low bitrate all make the model's job harder and leave vocal tails in the instrumental.

What goes in and what comes out?

It accepts audio or video files (MP3, WAV, M4A, AAC, OGG, FLAC, MP4, MOV, WEBM) within the size ceiling printed under the upload box, or a YouTube, TikTok or SoundCloud link. Each layer comes back as its own MP3 file, downloaded separately. The request and its results are saved to your activity history if you are signed in, and are never made public.

How does this compare to installing software?

All three routes separate audio; they differ in what you have to give up. Installed stem-separation software runs offline, has no upload limit and can batch whole folders, but it needs installing, usually wants a discrete GPU to be quick, and often has a command-line step that stops newcomers. Online tools usually upload the file to their own server to process it, so there is a size ceiling set by server memory. This page is different from both: the engine running today lives right on your device, so nothing is uploaded anywhere, and the size ceiling here is your own browser memory rather than a server's. The trade-off is that the on-device engine only supports 2 or 4 stems; the 5-stem option needs the server engine (spleeter) on a paid run, and the tool says so plainly whenever you pick an option that needs it, instead of handing back an empty file.

How do you check how clean a separation is?

The quick check is to listen closely to the loudest chorus, where instruments pile on top of the vocal, because that is where the model makes most of its mistakes. The reliable check is to open the stem in a spectrogram viewer and look at the frequency range: leftover vocal in the instrumental shows up as faint smears between roughly 200 Hz and 4 kHz, where a human voice concentrates its energy. Tách Nhạc has a spectrogram viewer that runs entirely in the browser, so you can drop the stem in and look straight away with nothing to install.

Who uses music separation?

Karaoke and cover singers who need a clean backing track of the exact original arrangement
Voice teachers who want students to hear the reference vocal line on its own
People learning production who want each instrument isolated to study the arrangement
DJs and mashup makers who need an acapella or a standalone instrumental
Video editors who need the lyrics gone but the music kept as a bed

Separating a song takes four steps, with nothing to install and no account required.

Pick your source
Drag a file from your computer into the upload box, or switch to the link tab and paste a YouTube, TikTok or SoundCloud URL. The length and size ceilings for one file are printed around the upload box.
Choose how many layers
Pick 2 stems if you only need the vocal and the backing track. Pick 4 stems to get drums, bass and instruments as separate files. 5 stems needs the server engine on the paid lane.
Hit Separate now
By default the AI works right in your browser through the engine already on your device, with no file uploaded anywhere — or pick the server engine (spleeter) in the engine panel to have the server process it instead. A progress bar appears on the page while it works.
Download each layer
Every finished layer gets its own download button right here on this page. Sign in before separating and both the request and its results are kept in your activity history.

By default the engine runs 100% on your own computer/phone (local, nothing is sent anywhere); pick the server engine (spleeter) in the engine panel if you would rather the server process it.
Quality depends on the source: a heavily compressed file or a dense arrangement will leave vocal tails in the instrumental.
An AI-separated backing track is not a replacement for one recorded separately in a studio. It is good enough for karaoke and practice, but a careful listener can still hear the processing.

These specifications describe the separation tool as it stands today. The real figure behind each limit is shown inside the tool itself, read from the engine manifest rather than written into this page.

Input and output

Accepted formats: Audio: MP3, WAV, M4A, AAC, OGG, OGA, OPUS, FLAC, WMA, AIFF, AIF, AMR Video: MP4, MOV, WEBM, AVI, FLV, MKV, M4V, WMV, 3GP (the audio track is taken out and separated)
Input sources: File upload, or a YouTube, TikTok or SoundCloud link
Output: Each separated layer is its own MP3 file, downloaded separately

Processing

Where it runs: On your own device (htdv4), or on the server (spleeter) — pick either in the engine panel The moises engine is still being built and not open yet
Stem counts: Pick 2 or 4 layers 5 layers needs the server engine on the paid lane; the free lane tops out at 2-4 layers
Engine status: The on-device engine (htdv4) and the server engine (spleeter) are both running for real. The moises engine is not open yet

Limits

Maximum length: On-device engine (htdv4): 6 minutes on a free run, 20 minutes on the paid tier The server engine (spleeter) applies its own tier-based ceiling, shown in the engine picker inside the tool
File size: On-device engine (htdv4): up to 100MB, bounded by your own browser's memory rather than by account tier The drop zone accepts more, because the server engine (spleeter) can take heavier files; whichever engine you pick is the ceiling that applies

Price and account

Price: Free within a weekly run budget; the tool shows how many runs you have left. Past that it is billed per minute per stem — the exact rate differs between the on-device engine (htdv4) and the server engine (spleeter), shown in full on the consent screen before anything is charged, never typed here
Account required: No
Privacy: Your files are never made public, and neither is your IP

Engine

Description: Pros: faster than HTDv4. Cons: separation is not as clean as HTDv4.
Runtime: Server
Guest: Free · 5 min/file · 2 stems
Signed in: Free · 7 min/file · 4 stems
Paid: 1E per minute per stem · 20 min/file

Description: Pros: separates many times cleaner than Spleeter. Cons: runs in your own browser, so it is slower and a weak device may not run it at all.
Runtime: On your device
Guest / Signed in: Free · 6 min/file · 100MB
Paid: 2E per minute per stem · 20 min/file

Description: The third-party Moises separation API.
Runtime: Server
Paid: 6E per minute · 20 min/file
Status: Coming soon

Sep 1, 2026
Unsigned Vietnamese variants in body and schema
The intro and page schema now include unsigned Vietnamese query forms (tach beat, tach giong hat, tach nhac va loi). The H1 stays accented Vietnamese.
Aug 26, 2026
The on-device engine (HTDv4) now outputs MP3
Every layer from HTDv4 now comes out as MP3 (256kbps) instead of WAV: a much smaller, faster download with no audible difference for most karaoke/practice use.
Aug 24, 2026
Plainer wording in the specifications table
The duration note in the Specifications tab and step 1 of the Instruction tab now read as a plain maximum instead of a finance-flavoured word. The figures and their meaning are unchanged.
Aug 24, 2026
Shorter engine names, descriptions now state pros and cons
Engine names shortened to Spleeter, HTDv4 and Moises for a tidier tab strip. Each engine description now states its pros and cons so you can compare at a glance. A server engine also shows a connection dot: green when the server is reachable, red when it is not.
Aug 23, 2026
Spleeter is now the default engine; plainer wording
The engine picker now defaults to Spleeter (server) instead of the on-device engine; if the Spleeter server is having trouble, the Separate button is temporarily disabled with a warning suggesting another engine. Replaced the bare English "local" in the Vietnamese notes with a plain-language phrase.
Aug 2, 2026
Server engine (spleeter) is live; player reuses one shared file
The server-side separation engine (spleeter) is now live. The engine panel offers two working choices: on your own device (htdv4) or on the server (spleeter), both metered and charged normally. In the link box: the downloaded file is fetched exactly once and reused for the preview player, the Download button, and both Tách paths (the on-device one uses the file directly, the server one reuses that same file's public url instead of fetching it again).
Aug 2, 2026
The link box goes one-tap: pasting auto-downloads, Tách is its own button
Pasting a link now auto-starts the download immediately (no separate click), reporting two honest phases: "Checking…" then "Downloading…". Once it finishes, the page shows a preview along with a Download button and a Separate button side by side. Downloading never waits on separating, and pressing Separate reuses the file that was already fetched instead of fetching it again. The two actions now spend two independent budgets, replacing the earlier merged-into-one-run flow.
Aug 1, 2026
The link box: a preview, and a download-only button
Pasting a link now shows a browser-only preview (no server call) when the platform is recognized, plus a separate "Just download (no separation)" button next to the separate button. Take the raw audio, separate it, or both, neither forces the other. A link-sourced separation request now goes through the real two-budget reserve flow (fetch + separate) instead of always falling straight into the pending stub as before; the visible outcome is unchanged since both sides of the pipe are still being built.
Jul 31, 2026
Local engine renamed to HTDemucs v4
The on-device engine's technical id changed from "separate-local" to "htdv4" (owner: the actual model is HTDemucs = Hybrid Transformer Demucs = Demucs v4, confirmed via the StemSplitio/htdemucs-onnx weights file). This is an identity rename, not a processing change: same model, same limits, same free-turn budget.
Jul 30, 2026
Copy corrected to match the engine that actually runs
The whole page (Intro, Instruction, Specifications tabs) was corrected to match the fact that the on-device engine (separate-local) has been running for real since 2026-07-29: processing happens in the browser rather than on a server, only 2 or 4 stems work today (5 stems needs the server engine, still being built), and the duration/size ceilings are read straight from the manifest instead of typed by hand.
Jul 26, 2026
Specifications moved to their own tab
The specifications and limits table left the middle of the Intro tab and now has its own Specifications tab, grouped into input and output, processing, limits, and price and account. The manual-service offer also moved above the tab strip and was rewritten around the separation cases a model cannot finish.
Jul 25, 2026
Now has its own page at /vocal-remover/
Music separation used to live only on the homepage. It now has its own tool page on the same frame as every other tool: hero band plus Intro, Instruction and Changelog tabs. The homepage still embeds the same tool. The Intro tab was also rewritten longer: a definition section, a specifications and limits table, and a comparison with installing software.
Jul 14, 2026
Separation tool opened to users
The first release of Tách Nhạc shipped with the separation tool: upload a file or paste a link, choose 2, 4 or 5 stems, and save the separation request to your activity history when signed in.

Vocal remover online

Remove vocals, pull karaoke beats and isolate stems with AI

Still hearing vocals in the instrumental?

What is music separation?

What is a karaoke backing track?

Should you pick 2, 4 or 5 stems?

Why does one song separate cleanly and another not?

What goes in and what comes out?

How does this compare to installing software?

How do you check how clean a separation is?

Who uses music separation?

Pick your source

Choose how many layers

Hit Separate now

Download each layer

Input and output

Processing

Limits

Price and account

Engine

Unsigned Vietnamese variants in body and schema

The on-device engine (HTDv4) now outputs MP3

Plainer wording in the specifications table

Shorter engine names, descriptions now state pros and cons

Spleeter is now the default engine; plainer wording

Server engine (spleeter) is live; player reuses one shared file

The link box goes one-tap: pasting auto-downloads, Tách is its own button

The link box: a preview, and a download-only button

Local engine renamed to HTDemucs v4

Copy corrected to match the engine that actually runs

Specifications moved to their own tab

Now has its own page at /vocal-remover/

Separation tool opened to users

More to read

Vocal removal for karaoke: a complete guide

Extract instrumentals for covers: studio-ready

Remove noise from phone recordings, fast

Getting a karaoke beat from any song: how it works

Removing vocals from a song: how-to and real limitations

How to remove drums from a song, and what it cannot do

How to remove bass from a song, or isolate the bass line

Related tools