State the test before you touch the file
Turning a song to instrumental feels like a single job until you actually try it. Then it splits into a dozen smaller questions: Which file should I use? Should I cut the intro first? What happens if I push the vocal down too far? Most guides answer these with features. This one answers them with a test.
Here is the test in one sentence: pick one short passage, run it twice with exactly one change between the two runs, then keep the version that survives your decision rule.
That is the whole method. Everything below is how to set it up and what to listen for.
Your decision rule
Write it down before you press anything, because your ears get tired and your standards drift. A workable rule looks like this:
I will keep the version where I can sing over the track for 30 seconds without hearing a second voice fight mine, and where the drums and bass still sound like drums and bass.
Two conditions, both required. A track that buries the vocal but smears the cymbals fails. A track that keeps everything crisp but leaves a ghost of the original singer fails too. If you cannot state a pass and a fail, you are not testing — you are auditioning.
1. Test setup
Choose a passage between 20 and 45 seconds. Long enough to contain a chorus, short enough that you can listen five times without your attention sliding off.
Pick the passage that scares you most, not the easiest one. A verse with a quiet vocal sitting on a soft pad will separate more cleanly than a chorus where the singer belts over stacked guitars. If your goal is a backing track for a live performance, test the part you will actually perform over.
Prepare the file
The source matters more than any setting you will touch later. Use the highest-quality copy you have. A compressed stream ripped from a video will carry artifacts that no separation mode can undo, because those artifacts are already baked into the waveform.
ShengXi Vocal Remover accepts videos from your Photos library and local audio files, so if your only copy is a video, you can pull it in directly rather than hunting for a separate audio export first. If a clean local audio file already exists on your device, use that.
Set a listening baseline
Before separating anything, play the original passage once at a fixed volume. Note three things in writing:
- Where the lead vocal sits — centered, doubled, drenched in reverb?
- Which instrument carries the low end?
- Any moment where the vocal overlaps a lead instrument, like the last line before a chorus.
This baseline is your reference. Without it, you will judge the result against a memory, and memories are generous.
Confirm your rights
You need permission to process and publish whatever you are working with. The app separates audio; it does not grant you rights to music or video. For a private rehearsal file that never leaves your device, this is usually a non-issue. For anything you plan to post or perform publicly, check the license first. Do that now, not after you have spent an hour on a track you cannot use.
2. First attempt
Run the passage once with a sensible default. Do not tweak. Do not preview five times first. The point of attempt one is to produce a reference, not a winner.
ShengXi Vocal Remover offers a fast mode and a fine mode. Start with fast. It gives you a result quickly, which keeps this stage cheap — you are going to listen, take notes, and change something anyway.
Export the separated accompaniment. Keep the vocal layer too, even if your plan is to discard it. You may want it later for comparison, and generating it twice costs you nothing right now.
What a first attempt usually gives you
Expect something usable but imperfect. Depending on the source, you will likely notice one of these:
- The lead vocal is largely gone, but a reverb tail or a doubled harmony survives underneath.
- Percussion sounds slightly watery, especially cymbals and hi-hats.
- The bass has thinned out, or gained a faint wobble on sustained notes.
- Backing vocals are reduced but not removed — they are mixed differently from the lead and separate less cleanly.
None of these means you did something wrong. Separation quality follows source quality, mixing choices, noise, and compression. A dense master with heavy limiting is simply harder to pull apart than a sparse arrangement. Do not promise yourself an identical copy of the original minus one voice. It does not exist.
3. What to observe
Now listen properly. Three passes, three different questions.
Pass one: the ghost check
Play the accompaniment and sing along at full volume. Can you hear another voice underneath yours? Not a hint of reverb — an actual vocal. If you can, note roughly where in the passage it appears. Precision matters: "around 0:18, right on the held note" tells you more than "the chorus is bad."
Pass two: the instrument check
Ignore the vocal problem entirely. Listen only to the instruments. Do the drums still have attack? Does the low end hold together? Is there a hollow, phasey quality on any sustained sound? Write down the single worst-sounding instrument. It will usually be one, not several.
Pass three: the sing-along check
Put the accompaniment on and perform the passage as you intend to use it. This is the only pass that matters for the decision rule. If you stumble because the track pulled your attention to a leftover vocal, that counts against it — even if the file sounded fine in passive listening.
Score the attempt
You now have two notes: the worst vocal leak and the worst instrument artifact. That is your baseline. Attempt two has to beat it under your own stated rule, and you decide that by listening, not by assumption.
4. Change one variable
This is where most people go wrong. They switch modes, trim the intro, adjust the vocal level, and re-export all at once. If the second version is better, they have no idea why. If it is worse, they have no idea which change caused it.
Change exactly one thing. Pick from this list:
- Processing mode. Re-run the same passage in fine mode instead of fast.
- The source file. Swap the compressed version for a higher-quality copy of the same passage.
- The passage boundaries. Start the clip a second earlier or later so the separation does not begin mid-phrase.
- The level balance. Import the result and use the separate vocal and accompaniment volume controls to reduce a surviving vocal layer before exporting.
For a first test, option one is the cleanest. It isolates processing quality and nothing else.
One note on mixing after separation
If your leftover vocal is faint rather than obvious, pulling the vocal layer down further is a legitimate fix, and it is often faster than re-running the whole passage. Preview the mix before you export. Then listen to the export itself, not just the preview — they are the same audio, but you will hear it differently once it is a finished file.
5. Record the result
Listen to attempt two with the same three passes. Then fill in a short table. On paper or in a note, it does not matter.
- Version: fast mode / fine mode / other change
- Worst vocal leak: timestamp and one-word description
- Worst instrument artifact: which instrument, what it sounds like
- Passes my decision rule? yes or no
- Would I perform over this? yes or no
The last two lines are the only ones that decide anything. Everything above them is evidence.
Read the outcome honestly
Three outcomes are possible, and only one of them means you are finished.
The second version passes. Use it. Keep the first attempt anyway — not because it might be better, but because if you revisit the track later with different ears, you will want to compare rather than start from scratch.
The second version fails the same way. The variable you changed was not the bottleneck. The problem is probably in the source: heavy compression, dense backing vocals, or an arrangement where the vocal and instruments occupy the same frequency range. Try a different passage from the same song before you try more settings.
The second version fails differently. You traded one artifact for another. This is informative rather than discouraging — it tells you the tradeoff exists and you get to choose which side of it you prefer.
Keep both attempts
Store both exports side by side and label them with the variable you changed, not with "good" and "bad." Labels like "fast mode" and "fine mode" still mean something in a month. Labels like "version 2" do not.
The result history inside ShengXi Vocal Remover keeps recent extractions around, which is convenient for short comparisons. For anything you intend to revisit, export it out. A file you can find is worth more than a result you have to regenerate.
Extending the test without breaking it
Once one passage passes your rule, extend to a second section of the same song. Do not change anything else. If the chorus passes but the bridge does not, you have learned something specific about that song rather than something vague about the tool.
For singers building a rehearsal set, the same method scales. Run one verse of each song, apply the identical decision rule, and note which tracks need a second pass and which do not. You will end up with a shortlist based on evidence instead of an afternoon of blind re-runs.
If you are working entirely on an iPhone, the workflow stays the same: import from Photos or local files, separate on the device, preview the mix, export, and file both attempts. The test does not change because the screen is smaller — only your patience for repeated listening does.
The one habit that makes this work
Write the decision rule down before you listen. It is the difference between a repeatable method and an afternoon of guessing. Everything else — modes, levels, exports, file quality — is a variable you can move one at a time.
Run the test once. Then run it again with one thing different. Keep both files, and let your own ears, measured against a rule you set in advance, make the call.