How to Isolate Instruments From a Song

Updated June 2026

To isolate an instrument from a song, run the file through an AI stem splitter: it separates the mix into vocals, drums, bass, and instrumental as individual audio files, and you keep the one you want. You can hear a free 30-second preview of every stem in your browser with no sign-up.

How AI source separation works, in plain words

A finished song is one stereo file — the parts are not stored separately inside it, so there is nothing to simply “un-mix.” What an AI separator does instead is recognize. The model behind Trackpeel was trained on a large number of songs where it could hear both the full mix and the isolated parts, so it learned what a human voice, a drum kit, and a bass line each look like in a spectrogram — the picture of which frequencies are playing at which moment. Given a new song, it estimates, moment by moment and frequency by frequency, how much of the sound belongs to each part, and rebuilds each one as its own audio file.

Those rebuilt files are called stems. Trackpeel returns four: vocals, drums, bass, and instrumental (everything that is not vocals, drums, or bass — guitars, keys, synths, and strings live together in this one). From those four it also derives the two most-requested combinations: the full instrumental with no vocals, and the acapella.

Which instruments can be isolated?

Being honest about this saves you time:

Drums — yes, as a dedicated stem. Kick, snare, hats, toms, and fills come out together as one percussion track.
Bass — yes, as a dedicated stem.
Vocals — yes, both as an isolated acapella and as its inverse, the instrumental.
Guitar, piano, synths, strings — not as separate stems. They stay together inside the instrumental stem. That stem is still useful for learning parts: with the vocals, drums, and bass stripped away, a guitar line or piano voicing that was buried in the mix becomes much easier to hear.

Step by step: isolate an instrument with Trackpeel

Upload the song. Drop in an MP3, WAV, M4A, or FLAC — or an MP4 or MOV video, and the audio is extracted for you. There is no app to install and no sign-up to start.
Let the model separate it. The split runs on a GPU and usually takes about a minute for a full-length song.
Preview every stem free. Each stem gets a 30-second preview at the same quality as the paid output, so you judge the real result — solo the drums, the bass, or the instrumental and listen before paying anything.
Download what you need. A free account lets you download each 30-second preview as an MP3 sample and keeps your splits in a Library. Full-length downloads are lossless WAV plus a 320 kbps MP3 per stem, from a $5 Day Pass — no subscription required.

When AI separation struggles

Source separation is very good in 2026, but it is not magic, and knowing the hard cases helps you set expectations before you commit to a download:

Dense, loud masters. Heavily limited or distorted mixes — where every element fights for the same frequencies — leave the model less room to tell parts apart, so stems can carry faint traces of their neighbors.

Live recordings. Room reverb and microphone bleed smear the boundaries between instruments.

Low-bitrate sources. A heavily compressed MP3 has already thrown away detail the model would otherwise use. Feed it the best copy you have — lossless if possible.

Unusual instruments. Sounds that sit between categories (a distorted bass synth, a vocal used as a texture) can land in the “wrong” stem or be shared across two.

This is exactly why the free preview exists: listen to the actual separation of your actual song first.

Honest alternatives: EQ and phase tricks

Before AI separation, isolation meant signal-processing tricks, and they are still free in any DAW — with real limits:

EQ carving. Low-pass a song around 100–150 Hz and you keep mostly bass and kick; band-pass the midrange and the vocal gets louder relative to the rest. The limit: instruments overlap heavily in frequency, so you always keep a blurry blend, never a clean single instrument. Fine for a rough practice reference, not for a usable stem.

Phase cancellation (the “center cut”). Invert one channel of a stereo file and everything panned dead center cancels out. It can crudely remove a centered lead vocal, but it also removes the bass and kick (also centered), leaves stereo reverb tails behind, and cannot isolate anything — it can only subtract the center.

If those trade-offs are acceptable for your use, they cost nothing. When you need an instrument genuinely on its own — drums clean enough to chop, a bass line you can transcribe — model-based separation is the approach that actually does it.

Frequently asked questions

Can AI isolate guitar or piano from a song?

Not as dedicated stems on Trackpeel. The model separates four stems — vocals, drums, bass, and instrumental — so guitar and piano stay inside the instrumental stem. Removing the vocals, drums, and bass still makes those parts far easier to hear and learn, but you will not get a guitar-only or piano-only file.

How do I isolate one instrument from a song for free?

Upload the song to Trackpeel and preview each stem — vocals, drums, bass, instrumental — free for 30 seconds with no sign-up. A free account also downloads those previews as 30-second MP3 samples. Full-length lossless WAV downloads are paid, from a $5 Day Pass.

How long does instrument isolation take?

About a minute for a typical full-length song. The separation runs on a GPU in the cloud, so it does not depend on how fast your own computer is, and nothing needs to be installed.

Is it legal to isolate instruments from a song?

Separating a track for personal practice, learning, or private study is generally fine. The recording still belongs to its rights holders, though — releasing, sampling, or monetizing an isolated stem requires permission or a license, the same as sampling the original record.

What file formats work?

Audio uploads: MP3, WAV, M4A, and FLAC. Video uploads: MP4 and MOV — the soundtrack is extracted automatically. Downloads come back as lossless WAV plus a 320 kbps MP3 for every stem.