A free vocal extractor and stem splitter that runs in your browser
Use this vocal extractor to pull the singing out of a finished song: an acapella for a remix or mash-up, a guide vocal to learn harmonies from, or a cleaner voice for transcribing lyrics. Switch on “Also keep drums, bass and other” and the same run becomes a four-way stem splitter for practice or sampling.
The AI is HTDemucs, Meta’s open music-separation model, running on your own computer through WebAssembly or WebGPU. No account and no upload: the song stays in your browser and only the model is downloaded, once. A classic centre-extraction mode needs no download at all.
How to use the Vocal Extractor
- Drop a song on the upload box or click to choose one — MP3, WAV, FLAC, M4A, OGG, Opus or a video file.
- Leave “AI separation” selected and tick “Also keep drums, bass and other” if you want every stem, not just the voice.
- Press “Extract vocals” and confirm the one-time 166 MB model download. Choose “First 30 seconds (quick test)” to preview on a slow computer.
- Compare the original with “Vocals (acapella)” and the other stems in “Listen and compare”; the player stays at the same point when you switch.
- Choose a format under “Save as” and press “Download vocals”, “Download instrumental” or “Download all stems (ZIP)”.
Vocal Extractor features
Four stems from one run
Vocals, drums, bass and other instruments, plus a ready-made instrumental (drums + bass + other).
Meta’s HTDemucs model
The hybrid transformer version of Demucs, released under the MIT licence, works on the waveform and the spectrogram at the same time.
Stems that add up
The four parts sum back to the original mix, so remixes line up sample-for-sample with the source.
ZIP of every stem
Save all five files in MP3, WAV or FLAC in one download, named after your song.
Classic centre extraction
No download: keeps only what is identical in both channels, ignoring bass below 120 Hz and treble above 12 kHz by default.
Nothing leaves your device
Separation runs in a background worker in your browser; the page never receives your audio.
What each stem contains
HTDemucs was trained on four sources, so every sound in the song ends up in one of them.
| Stem | What you hear in it | Typical use |
|---|---|---|
| Vocals (acapella) | Lead and backing vocals, spoken parts, some breaths and reverb | Remixes, mash-ups, learning harmonies, lyric transcription |
| Drums | Kit, claps, shakers and most percussion | Drum practice, sampling, tempo checks |
| Bass | Bass guitar, synth bass, low 808s | Learning bass lines, re-amping |
| Other | Guitars, piano, synths, strings, brass — everything else | Practising your instrument, isolating a hook |
| Instrumental | Drums + bass + other | Karaoke and backing tracks |
How the vocals are isolated
The song is resampled to 44.1 kHz stereo and processed in 7.8-second windows that overlap by 25%, with a linear cross-fade where they meet. In each window HTDemucs runs a time-domain branch on the waveform and a frequency-domain branch on the spectrogram; a cross-domain transformer connects them before both are decoded into four sources. The page then adds the two branches, applies the inverse spectrogram and stitches the windows together.
Model: HTDemucs from Meta’s Demucs project (MIT), converted to ONNX as Ghilda/htdemucs-onnx (MIT), 166 MB. Pre- and post-processing: demucs-web 1.0.2 (MIT). Runtime: ONNX Runtime Web 1.30.0 (MIT), using WebGPU where available — if the stems of the first loud window do not add up to the input, the tool restarts on the processor. Measured in Chrome on a mid-range laptop: about 4.6 s per window on integrated graphics and 20 s on the processor, with roughly 3 GB of memory in use.
Classic centre extraction
Without the AI model, the tool can still keep the centre of a stereo mix, where most lead vocals sit. For each slice of the spectrum it compares left and right and keeps only what is identical in both; anything panned, widened or out of phase is dropped. In this page’s automatic test, tones panned left, right or wide fall by about 60 dB while the centred tone stays within 0.1 dB.
The catch is that bass, kick and snare also live in the centre. “Ignore bass below” and “Ignore treble above” keep most of them out, but any other centred sound comes through with the voice — fine for checking lyrics or a melody; use AI for anything you will publish.
Getting the most from an acapella
- Find the song’s tempo with the BPM Detector and its key with the Key Finder before you build a remix around the vocal.
- Need the music without the singer? The Remove Vocals page is the same engine set up for karaoke tracks.
- Reduce hiss in a quiet acapella with the Audio Noise Remover — save as WAV or FLAC first.
- Writing down lyrics? A clean acapella helps Audio to Text, although singing is always harder to transcribe than speech.
- Rights still apply: an extracted vocal belongs to the artist and label. Use it for personal practice, study, or remixes you are licensed to make.
Vocal Extractor FAQ
How do I extract vocals from a song?
Drop the song here, press “Extract vocals”, confirm the model download, and wait for the progress bar. Then press “Download vocals” to save the acapella as MP3, WAV or FLAC.
Can I split a song into drums, bass and other stems?
Yes. Tick “Also keep drums, bass and other” before you start; when the separation finishes, “Download all stems (ZIP)” saves vocals, drums, bass, other and the instrumental together.
How good is the acapella?
HTDemucs gives a usable acapella on most studio recordings. Expect small artefacts — a hint of hi-hat in breaths, or reverb tails that end abruptly — in loud, dense mixes. Live recordings with crowd noise are the hardest.
Why is the model 166 MB?
It holds the full HTDemucs network in 32-bit precision, matching Meta’s original model. It downloads once and your browser keeps it, so later songs start in seconds.
How long can the song be?
Up to 10 minutes for AI separation (use “First 30 seconds” to preview longer files) and 20 minutes in classic mode.
Is it safe to use with unreleased music?
Yes. Separation happens entirely in your browser, so demos and unreleased mixes stay private.
Can I use an extracted vocal in my own release?
Only with permission from the copyright owners of the original recording and song. Separating a track does not change who owns it. This is general information, not legal advice.