Vocal Remover & Karaoke Maker

What is vocal separation?

Vocal separation (also called vocal removal or stem separation) is the process of splitting a finished song into two parts: the vocals — the singing, rap or speech — and the instrumental (everything else: drums, bass, guitars, keyboards, synths). It is the technology that makes karaoke possible without hiring a live band, and it lets producers sample, remix or cover songs that were never released as separate tracks.

Until the late 2010s this required either access to the original studio multi-tracks (rare) or phase-cancellation tricks that only worked on mono recordings. Modern AI models such as demucs take a fully mixed song as input and produce two clean stems, even on tracks with heavy reverb, side-chain compression or layered backing vocals.

How to use this tool

1. Prepare your audio. Any common format works — MP3, WAV, FLAC, OGG, M4A, AAC, OPUS, WMA, AIFF or M4B. Sample rate is normalised automatically; bit depth is preserved.
2. Drop the file into the upload box above, or click to browse. The selected file name will appear below the box.
3. Pick an output format. Choose MP3 for a small file you can play on any device, or WAV for an uncompressed master suitable for further editing in a DAW.
4. Click Separate. The browser uploads the file to our servers, where the AI model (htdemucs_ft) processes it. The progress bar shows a "loading the AI model" step the very first time you use the tool, then "separating vocals and instrumental" while the song is being analysed. A 3-minute song typically takes 60–180 seconds.
5. Download the ZIP. When the job finishes, click Download ZIP (vocals + instrumental) to save a single archive containing the two stems. Both stems share the source file's tempo and pitch — they are ready to drop straight into a project.

Tips for the best results

• Use the highest-quality source you have. Lossy formats like 128 kbps MP3 introduce artefacts that the AI cannot undo. A 320 kbps MP3, AAC, FLAC or WAV source will always sound noticeably cleaner.
• Shorter clips separate faster and better. If you only need a chorus or a verse, cut the relevant region in your DAW first; the model analyses the whole file.
• Mono is fine. Most pop, rock and hip-hop releases are mixed in stereo, but the AI handles mono sources just as well.
• Heavy mastering is your enemy. Loudness-war tracks that have been brick-wall-limited strip the spatial cues the model relies on. A song that has been professionally mixed but not mastered usually separates much better than the same song after -8 LUFS limiting.

What "instrumental" actually means here

demucs htdemucs_ft is a hybrid transformer that splits a song into four internal sources — vocals, drums, bass and other (guitars, keys, synths, anything that is not vocals, drums or bass). The --two-stems vocals mode this tool uses returns two stems:

• Vocals — the model's vocal prediction.
• Instrumental — the original song minus the vocal prediction. This is computed by subtracting the vocals from the input, which is what makes the instrumental sound "hollow" in some frequencies. That is normal, and is the same trade-off that every AI vocal remover makes.

If you need finer-grained stems (drums only, bass only, etc.) the full demucs model exposes them as a separate option; the two-stem mode is what most karaoke and remix workflows actually need.

Common uses

• Karaoke backing tracks — drop the instrumental into your karaoke app or DJ software.
• Acapella extraction for mashups, remixes and bootlegs.
• Sampling — pull a clean vocal or instrumental loop out of a finished record without paying for an official remix clearance.
• Practice and teaching — sing along with the instrumental, then compare your cover to the original acapella.
• Audio restoration — isolate vocals for cleanup, noise reduction or pitch correction without touching the backing track.

Technical notes

The model runs CPU-only on our server. The first request after a fresh container start spends about 20–30 seconds loading the 340 MB model weights into memory; subsequent requests start processing immediately. Audio larger than the server's per-job limit (currently 50 MB) will be rejected — split long files first if needed. There is no time