JamScore AI

Home / Guides / How to Isolate Instrument Stems from a Song

How to Isolate Instrument Stems from a Song

Stem separation lets you pull a song apart into its individual instruments, drums, bass, guitar, piano, vocals, without a multitrack session or the original project files. Here is how to do it, what the results sound like, and where the technique genuinely falls short.

How to isolate instrument stems from a song with AI stem separation

Last updated August 2026 ยท 7 min read

What stem separation actually does

When a record is mixed and mastered, all the instruments are combined into a single stereo file. Stem separation attempts to reverse that process, to unmix the mix. It is a hard problem, and for a long time the results were poor enough that serious musicians ignored it entirely.

The situation changed as transformer-based models were applied to the task. JamScore AI uses dedicated band-split deep neural architectures trained specifically for music source separation. It processes the frequency content of your track and assigns each time-frequency bin to the most likely source: vocal, drum, bass, guitar, piano, or a catch-all "other" category. The result is six independent audio streams, each one containing predominantly (though not exclusively) that instrument.

What you do with those streams is up to you, listen and study, export for remixing, loop for practice, or feed them into a transcription tool.

Step-by-step: isolating stems in JamScore AI

  1. Upload a high-quality audio file

    Drag an MP3, WAV, M4A, or OGG file onto the upload area (or click to browse). Files up to 10 minutes long are supported. Use the highest-quality version of the track you have, a WAV from a purchased download will give the model more spectral detail to work with than a heavily compressed stream rip. Six-stem separation requires a Pro or Studio plan; the free tier produces vocals and instrumental only.

  2. Wait for the six-stem separation to complete

    The model runs locally in your browser using WebAssembly. Expect 30 seconds to two minutes depending on the track length and your device. A progress bar tracks the processing; the tab can be left in the background. Once it finishes, the studio mixer loads with six labelled channels: Vocals, Drums, Bass, Guitar, Piano, Other.

  3. Solo the stem you want to study

    Press the Solo button on a channel to mute all others and hear only that instrument. Press it again, or click Solo on a different channel, to switch focus. This is the fastest way to audition what the model has produced before you commit to exporting anything.

  4. Adjust EQ and volume to taste

    Each channel has a three-band EQ (low, mid, high) and a volume fader. If you are isolating a bass guitar stem and want to hear the low fundamental more clearly, boost the low band. If a guitar stem has an unpleasant boxy mid-range character from bleed, cut the mid. These are non-destructive, they only affect playback and export from this session.

  5. Use A/B loop markers to drill a section

    Set loop in and out points around a bar, a fill, or a specific chord change. The section will repeat continuously, which is useful for careful listening, transcription, or, if you are a vocalist, practising your own part over the backing.

  6. Export stems as WAV

    Click the download icon on each channel to export that stem as a lossless WAV file. You can export all six stems individually, then import them into any DAW (Ableton, Logic, Reaper, GarageBand) for further editing. There is no MP3 export, WAV only.

What each stem contains

Vocals

Lead and backing vocals, any vocal ad-libs, and sometimes close-mic'd spoken word or rap. Choir pads and vocoder parts usually appear here too, since their spectral shape resembles a human voice.

Drums

The full kit: kick, snare, hi-hats, toms, cymbals, and percussion. Drum machine patterns are generally well-separated. Handclaps and shakers sometimes split between Drums and Other depending on how they were produced.

Bass

Bass guitar (both DI and amp-miked), synth bass, and 808 sub patterns. In hip-hop productions with a very sub-heavy 808 that also contains pitched melody, the model may split part of it into the Piano channel, 808s with significant harmonic content look spectrally similar to a bass piano note.

Guitar

Electric and acoustic guitar. Heavily distorted rhythm guitars sometimes share spectral space with keyboards, particularly if both are mid-heavy. Expect some bleed in dense rock mixes with wall-of-guitars production.

Piano

Acoustic piano, electric piano, and many keyboard sounds. Clean-toned synth pads with slow attacks can end up here as well. The piano/guitar boundary is one of the trickier separations, particularly when both instruments occupy the upper-mid range simultaneously.

Other

Everything that does not fit cleanly into the five categories above: strings, brass, choir beds (non-vocal), synthesiser leads, sound design, effects, and any acoustic instrument the model cannot reliably classify. In orchestral or cinematic music, a large portion of the arrangement will land here.

Where the model struggles: Guitar and piano bleed into each other most noticeably in piano-driven pop ballads and jazz arrangements where the piano voicings cover the same register as the rhythm guitar. Dense rock mixes with layered guitars, bass, and keyboards all fighting for mid-range space are similarly difficult. Mono or vintage recordings are the hardest cases, without stereo-field information, the model relies solely on spectral characteristics, and artefacts (a phasey or "underwater" quality on isolated stems) are more audible. The de-reverb tool can help with heavily ambient recordings before you run separation.

Transcribing an isolated stem

One of the most practical uses of a clean stem is transcription. Once you have isolated the bass guitar, for example, you can send that stem through the MIDI transcription tool to get a playable MIDI part, standard notation, or guitar/bass TAB. The same applies to piano parts, drum patterns (drum score output), and melodic stems.

This is a far more reliable route to accurate transcription than feeding the full mixed audio into a transcription tool, the model has much less interference to deal with when the target instrument is already isolated. See the companion guide on converting a song to MIDI and sheet music for the full transcription workflow.

If you have printed sheet music you want to synchronise with your separated stems, the optical music recognition feature can convert a PDF or image scan to MIDI and MusicXML, useful for following along with the score while the isolated stem plays.

Separate your first song free

Three songs per month on the free plan, no account needed to get started. Pro unlocks unlimited six-stem separation for $9/month.

Open JamScore AI

Practical use cases

Frequently asked questions

Which stems does JamScore AI separate into?

The Pro and Studio plans use six-stem separation: vocals, drums, bass, guitar, piano, and other. The free plan produces two stems: vocals and instrumental.

Will the bass stem contain any guitar bleed?

Possibly, especially in dense arrangements. Guitar and bass share overlapping frequency ranges, and in tracks where both instruments are prominent and similarly EQ'd, the model will occasionally assign some guitar harmonics to the bass channel or vice versa. The effect is usually subtle.

Can I transcribe an isolated stem to MIDI or sheet music?

Yes. After separating, you can send any stem through the transcription tool to get MIDI, staff notation, guitar or bass TAB, or a drum score. This is particularly useful for learning basslines or piano parts from songs you cannot find tabs for.

What goes into the 'other' stem?

Anything the model cannot confidently assign to vocals, drums, bass, guitar, or piano. This typically includes strings, brass, choir pads, synthesiser leads, handclaps, and sound effects. It is essentially a catch-all for instruments outside the five named categories.

Does stem separation work on live recordings?

It works, but with lower accuracy than studio recordings. Live recordings tend to have significant room bleed between microphones, meaning drum overhead mics pick up guitar amplifiers, vocal mics pick up snare, and so on. The model does its best, but the bleed is baked into the audio and cannot be fully undone.

How long does separation take?

Typically 30 seconds to two minutes for a full song, depending on track length and your device's CPU. Processing runs in your browser via WebAssembly, so a faster machine will complete it more quickly. The tab can be left in the background.