There are exactly two ways to make an instrumental. From the original session, mute the vocals and bounce — that is perfect and takes two minutes. From a mixed file, use AI separation — that is an estimate, and it will leave some residue behind. Which route you take determines the quality ceiling, so it is worth checking whether the session exists before you reach for a tool.
- The original session, if it exists
- Otherwise, the highest-quality mixed file you have, plus a separation tool
- A DAW to check and repair
- Ten minutes for route one, an hour for route two
This is the correct method whenever it is available.
Not just the lead — backing vocals, ad libs, doubles, vocal chops, and anything on a vocal group.
This is where most session-based instrumentals go wrong:
- Sidechain compression keyed off the vocal. With the vocal muted, the ducking stops and the instrumental sounds different. Print the ducking or replace the trigger.
- Reverb and delay returns fed by the vocal. Mute those too, or the vocal reverb continues without the vocal.
- Bus compression on the master reacting to the vocal. The instrumental will now be less compressed and slightly louder. Re-check the level.
An arrangement written around a vocal has holes where the vocal was. Sometimes that is fine. Often the instrumental needs a lead instrument doubling the vocal melody in the chorus, or a longer pad, to stop it feeling empty.
Bouncing the same mix minus the vocal is not the same as mixing an instrumental. Bring the guitars and keys up, re-check the level, and master it as its own piece.
Lossless if you can. The separation model needs the high-frequency detail a low-bitrate MP3 has already thrown away.
Two-stem separation — vocals and instrumental — usually gives a cleaner instrumental than a 4-stem split summed back together, because there are fewer boundaries to get wrong.
Options include Demucs, Moises, lalal.ai, and Veena, which separates in the browser and drops the result straight onto a timeline for the repair work. Veena needs a desktop browser and a connection, so it is not an option on a phone.
| What you hear | Why | What helps |
|---|
| Ghost vocal in the reverb tails | Reverb belongs to the vocal but sounds diffuse | Low-level noise gating on the residue, or accept it |
| Dull, papery high end | Cymbals and sibilance overlap and got split badly | Gentle high shelf, or layer a fresh hi-hat |
| A hole where the vocal was | The model removed instrument energy along with the voice | Layer a pad or double the melody |
| Backing vocals left in | The model treated them as instruments | Try a different model, or accept them as texture |
Aggressive processing to chase the last trace of a vocal usually damages the instrumental more than the residue does. Get it clean, then stop.
A sparse, dynamic, well-recorded song can separate into an instrumental that passes as the real thing. A brickwalled modern pop master with layered harmonies and heavy reverb will not. Test on the loudest chorus, because that is where it fails.
Making an instrumental of somebody else's recording for private use is one thing. Releasing, distributing, or monetising it is another — you own your separation output as a file, but you own no rights in the underlying recording or composition. Instrumental versions of commercial songs need clearance from the same rights holders as any other use. Not legal advice, and it varies by country.
Related reading: how to split stems from a song, how stem separation works, and AI music and sample clearance.
Start making music in Veena
Free, browser-based, no downloads required.
Try Veena Free