Ever wished you could isolate a guitar riff for practice, or take out a vocal track for your very own home karaoke mix? Not long ago, that was the stuff of a high-end recording studio, the kind that costs a fortune, demands serious technical know-how, and needs more square footage than most of us will ever see in land-scarce Singapore. For the average music lover, it was a pure pipe dream.
Not anymore. AI-powered stem separation* uses artificial intelligence and deep learning models to "unmix" a finished stereo track into its individual parts. Vocals, drums, bass, and instruments, separated into their own tracks, all without ever touching the original studio session files. Think of it as mathematically reversing the recording and mixing process, a trick the industry calls "de-mixing" or "source separation."
What was once locked inside professional setups now sits beside us, from bedroom producers in Bishan to weekend buskers in Orchard. Here’s how this tech opened those studio doors
*For non-commercial use only. It is your responsibility to know and abide by copyright laws in your country, and your use may require permission from the copyright holder(s). Review terms associated with the use of prerecorded or licensed music and streaming services before using BandBox.
Let's start with the basics. In music, "stems" are the individual building blocks of a song, the separate tracks a producer layers together to create the final mix. Normally, these are locked away in private project files that only the original team can open.
When AI steps in, it typically breaks a song down into four main categories: vocals, drums, bass, and "other" (the catch-all for guitars, keyboards, and synths). Earlier tools could attempt this, but the results often sounded "watery" or metallic, like your track had taken a dip in the pool. Modern models such as Hybrid Transformer Demucs (HTDemucs) have changed the game, delivering far cleaner audio with fewer odd artefacts. Better still, the latest tech can split a full song in seconds, working right inside your browser or on your own gear.
Here's where it gets clever. Old-school tools tried to "cut" frequencies the way a basic EQ does, a bit like trying to remove the salt from your laksa with a sieve. Modern AI audio separation* uses trained neural networks instead, and the process roughly follows four steps:
The Training Phase: The AI listens to thousands of songs where every isolated track is already known. Over time, it learns to recognise the unique harmonic fingerprint of a snare, a bassline, or a human voice, whether it's pulling apart a Mandopop ballad or a homegrown indie track.
Detailed Audio Analysis: When you upload a new song, the AI converts the waveform into a 2D spectrogram, a visual map of how the frequencies shift over time.
Mask Estimation and Reconstruction: The system predicts which sounds belong together, then builds "masks," precise instructions that carve the audio into clean, separate outputs.
Customisable Muting: Most tools let you dial a part down by 25%, 50%, or 100%. No more hunting around for a "[song] without vocals" version, or running the same track through three different apps just to practise.
This is where the fun begins. Here's how musicians are putting the technology to work every day:
Music Production and Sampling: Producers can lift and clean a drum break or bassline from a vintage record, then build something completely new on top of it.
Remixing and Mashups: Want to flip a track? You can isolate a single vocal and craft your own remix, no need to beg to use the polytechnic’s sound studio after hours.
Music Education and Practice: Struggling with a tricky guitar solo? Isolate it and learn by ear. Or strip out the drums entirely and become the band's new percussionist for the night.
Audio Post-Production: Video editors can extract clean dialogue from a noisy stereo clip, saving hours of tedious editing for that next upload or wedding montage.
The ecosystem has exploded, and there's now something for every skill level.
For quick, no-fuss results, web-based platforms like LALAL.AI, Moises.AI, Loudly, and StemSplit let you separate a track straight from your browser, with no installation required.
If you're working at a more serious level, the tech is now baked into industry-standard software. You'll find it in Logic Pro (as Stem Splitter), Ableton Live, FL Studio, and iZotope RX 11, ready to slot neatly into your existing workflow.
As impressive as all this is, these are not the best JBL speakers ever, full stop. AI stem separation* is rarely perfect.
The most common gremlin is "bleed," a faint ghost of the drum kit or a guitar that still lingers on the vocal track. And while the tech gets sharper every iteration, heavily compressed or extremely busy mixes can still trip it up. The denser the song, the harder the AI has to work to tell every layer apart.
Here's the exciting part for performers, professionals, and people who just want to jam. The JBL BandBox Solo and JBL BandBox Trio are among the first portable amps to bring pro audio out of the studio and into your home audio system. Both run Stem AI Real-Time Separation* (with any track you perform live through it needing to be one you own), which builds the de-mixing magic right into the gear, so you can pipe the audio you want, and only the audio you want, straight from the speaker.
Built for the one-person show, the JBL BandBox Solo is your grab-and-go portable speaker. Drop the vocals out of a track to turn any song into your home karaoke, or strip back the instruments to drill a part on repeat.
When you're going for big sound, the JBL BandBox Trio steps up as the best for budding musicians, for a reason. Use its Stem AI separation* to "front the band" by muting the original vocals, or "take the solo" by pulling out a guitar line and playing it yourself.
Not long ago, unmixing a song was a job reserved for producers with the right software and a locked folder of session files. Today, the JBL BandBox range and the wider world of AI-powered stem separation* have handed that power to everyone, making it easier than ever to learn, remix, and perform on your own terms.
Ready to hear your music like never before? Explore the AI-powered JBL BandBox series at the JBL Online Store and start unmixing your world today.
*For non-commercial use only. It is your responsibility to know and abide by copyright laws in your country, and your use may require permission from the copyright holder(s). Review terms associated with the use of prerecorded or licensed music and streaming services before using BandBox.
