mvsep.com

What is MVSEP?

MVSEP is a free, AI-powered web tool for splitting audio into separate parts. Its main job is stem separation: you upload a track and it returns vocals and instrumental as separate files. It can also extract lyrics text from audio. The site lists a large menu of separation types covering vocals, drums, bass, piano, guitar, wind, strings, percussion, keyboards, MIDI and super-resolution, so it is aimed at people who need individual parts rather than a single mixed file.

Who it is for

• Singers and hobbyists who want a karaoke or backing track.
• Producers and remixers who need clean stems to edit or sample.
• Transcribers and lyric writers who want a vocal-only file to work from.
• Anyone testing how different separation models behave on the same song.

How it works in practice

You drag a file onto the page, pick a separation type, and the service processes it. Some presets are labelled "official" and use one algorithm pass; others are ensembles that combine several models. A few options are marked Premium only, and the page shows monthly usage counts and user ratings next to each preset, which is a useful signal when you are unsure which one to choose.

Choosing a preset

A practical next step

If you are new to it, start with a short excerpt of your track and try the Instrumental (vocal subtraction) preset, then compare it with a karaoke preset. If you need both a vocal and an instrumental file from one pass, look at the ensemble options. For lyric work, run a vocal-only separation first, then use the lyric extraction on that cleaner file. If you want a second opinion on separation quality generally, community discussions on Reddit and audio forums often compare tools, but MVSEP's own preset ratings and usage counts are the most direct guide to what works on this service.

How do I separate vocals and instrumental from a song using MVSEP?

Go to MVSEP, drag your song file onto the upload area, then pick a separation preset. The default choices are built around two different goals: pulling out a clean vocal, or rebuilding the instrumental.

Choosing a preset

• Instrumental (vocal subtraction) – the general-purpose choice. It rebuilds the backing track as the original mix minus the extracted vocal, so the result keeps the original phase. MVSEP's own listing says this gives cleaner transients than a model's direct instrumental output, and it is the most-used preset on the page.
• Karaoke (instrumental only) – removes lead and backing vocals, leaving a plain backing track to sing over.
• Karaoke pack (keep backing vocals) – strips only the lead vocal and keeps backing vocals in the instrumental, plus gives you the isolated lead vocal as a second file. Useful if you want the harmonies intact.
• Clean acapella – isolates the vocal and then removes background noise, aiming at a one-click clean vocal. Marked as Premium-only.
• Ensemble vocals + instrumental – runs several models and averages them, producing both a vocal and an instrumental stem in one pass. Premium-only.

Practical notes

If you are unsure, start with Instrumental (vocal subtraction) for a backing track and the standard vocal preset for an acapella. Presets that run multiple models or two passes take longer and, on MVSEP, some are restricted to the Premium tier.

The extracted vocal is not the same as a studio acapella: bleed from instruments and reverb tails will still be present, and dense or heavily compressed mixes separate less cleanly than sparse ones. Treat the output as a starting point for remixing, practising or transcription rather than a finished master.

As a next step, run the same song through two presets and compare them on the section you care about most — a chorus with stacked harmonies, for example — before committing to one for the whole track.

Which MVSEP preset should I choose for karaoke if I want to keep backing vocals?

Choose Karaoke pack (keep backing vocals). MVSEP describes it as stripping the lead vocal while keeping backing vocals in the instrumental, and it also returns the isolated lead vocal as a second file. That second file is the practical difference from the other karaoke option.

How the two karaoke presets differ

Pick Karaoke (instrumental only) when you want a plain backing track with no vocal traces at all — for example, a live performance where backing vocals would just clutter the mix. Pick the pack when the song's backing vocals are part of the arrangement (gospel-style harmonies, call-and-response, layered choruses) and you want them intact.

A practical next step: run one song through the pack, then listen specifically for two things — whether the lead vocal is fully gone, and whether any backing vocal bleeds into the separate lead-vocal file. If the backing vocals sound thin or phasey, try the Instrumental (vocal subtraction) preset instead; MVSEP notes it rebuilds the instrumental by inverting the extracted vocal and mixing it back, which preserves the original phase and tends to give cleaner transients than a model's direct instrumental output. It does not, however, distinguish lead from backing vocals, so it is a fallback rather than a replacement.

If you want to compare options side by side before committing, the preset list on MVSEP shows community usage counts and ratings under each entry, which is a reasonable signal for which separation types people actually trust on real material.

What are the differences between the free and Premium separation presets on MVSEP?

Free presets cover the core jobs: pulling out vocals, rebuilding an instrumental, or making a karaoke track. Premium presets are the more experimental or compute-heavy variants — ensembles, phase correction, de-noising — and they are the ones that can produce noticeably cleaner stems when a basic pass leaves artefacts.

What the free presets do

• Instrumental (vocal subtraction) — rebuilds the backing track as original-minus-vocals, so the result keeps the original mix's phase and tends to have cleaner transients than a model's direct instrumental output.
• Karaoke pack (keep backing vocals) — strips the lead vocal but keeps backing vocals, and also gives you the isolated lead vocal as a second file.
• Karaoke (instrumental only) — removes lead and backing vocals, leaving a plain backing track.
• Phase-corrected instrumental (Deux) — a single-model take on phase correction: one separation pass, then a spectral invert to rebuild the instrumental.

What Premium adds

• Clean acapella — isolates the vocal, then removes background noise in one click.
• Ensemble vocals + instrumental (124-bands weighted) — weighted average of three models, producing both a vocal and an instrumental stem from one pass per model.
• Phase-corrected instrumental — keeps the HyperACE v2 instrumental timbre but resynthesises it on a cleaner phase, reducing phase smearing.
• De-noised instrumental (double pass) — runs the vocal model twice and averages the results, cancelling model noise that isn't phase-locked to the input. This costs two algorithm passes, which is a good illustration of why these sit behind Premium.

How to choose

A practical next step: run one free preset first, listen for the specific problem (hiss, watery cymbals, phasey vocals), then pick the Premium preset that targets that problem rather than trying them all. If you only need a rough reference track for practice, the free options are usually enough; if the stem will be mixed into a release, the Premium variants are where the extra passes earn their keep.

How can I extract lyrics text from an audio file with MVSEP?

MVSEP offers a lyrics-extraction step alongside its vocal/instrumental separation, so you can go from a mixed song to a text transcript without leaving the site. The workflow is upload-first: drag your audio file onto the page, choose a separation preset, and let it process.

For lyrics specifically, the practical route is:

1. Upload the track to MVSEP (drag and drop onto the upload area).
2. Separate the vocals using a vocal-focused preset, since lyric extraction works far better on an isolated vocal stem than on a full mix with drums, bass and guitars competing for attention.
3. Run the lyrics extraction on that vocal output to get the text.

Why the vocal-first step matters: transcription tools mishear words when instruments mask consonants or when backing vocals overlap the lead. A clean acapella gives the recogniser a much better signal, which is the main reason to chain the two features rather than expecting one pass on the original file to work well.

Choosing a preset for lyric work

The page evidence shows a Premium tier exists (the "Clean acapella" and several ensemble presets carry a 🔒 Premium marker, with a plans page at /plans), while other presets are listed as basic. If you only need lyrics occasionally, start with a free vocal preset and check whether the transcript is accurate enough before considering paid options.

A realistic scenario

Say you have a live recording with crowd noise and a loud band. Running the full mix straight into lyric extraction usually produces garbled lines. Instead, separate the vocal first, then transcribe the isolated stem — the difference in accuracy is typically large, and it costs you one extra processing step rather than any new tool.

Next step: try one song end-to-end — separate vocals, then extract lyrics — and compare the transcript against the original. If accuracy is poor, that tells you the bottleneck is the vocal isolation quality, and a stronger preset (or a Premium one) is the lever to pull.

What is the difference between the official instrumental and phase-corrected instrumental presets?

The official Instrumental (vocal subtraction) preset rebuilds the backing track by taking the extracted vocal, inverting its polarity, and mixing it back with the original file. Because the result is built from the source mixture, it keeps the original phase and tends to preserve clean transients. The Phase-corrected instrumental presets instead keep a model's instrumental magnitude but replace its phase with a cleaner phase, either from a second model (the Premium version) or from a spectral invert of the mixture (the "Deux" version). The goal is to remove phase smearing while retaining the fuller timbre of the instrumental model.

In practice:

• Vocal subtraction is the safer, more literal choice. It sounds closest to "the original minus the voice," which suits remixing, sampling, or any case where you want the backing to sit in the same stereo image and timing as the source. Its weakness is that any vocal energy the model missed leaks back in, and artefacts in the extracted vocal can colour the result.
• Phase-corrected is for when the model's instrumental sounds good in tone but slightly blurry or smeared. Borrowing a cleaner phase can tighten the stereo image and definition. The trade-off is complexity: the Premium version runs two models, and the Deux version relies on one model plus a spectral invert, so results depend more on how well those components match.

A useful decision rule: if you plan to mix the instrumental with other stems or need it to align with the original, start with Instrumental (vocal subtraction). If the model's instrumental already sounds fuller but lacks focus, try a phase-corrected preset and compare the stereo image and transients against the subtraction result.

If you want to hear the difference on your own material, upload one track and run both presets, then solo the high frequencies and listen for smearing around cymbals and reverb tails. MVSEP lists both under its presets, with the phase-corrected variants marked Premium.