If you're wondering why anyone would want to pull a perfectly good song apart, I have a simple answer.
I'm a bass player. If I want to learn John McVie's bass line from Fleetwood Mac's Dreams, I'd rather hear that bass part on its own. And once I've learned it, I want the exact opposite: everything except the bass, so I can step into McVie's shoes and play along with the rest of the band.
Until quite recently, neither was particularly easy. For most of recorded music's history, the final mix really was final. Once the bass, drums, guitars and vocals had been mixed down to stereo, getting the original ingredients back out was rather like trying to recover the RAW file from a JPEG.
Pleasingly, that's no longer the case.
AudioPrism is a US$79.99 Windows application (currently $59.99) that promises to pull finished recordings apart on your own PC, turning them back into individual tracks for vocals, bass, drums and other instruments. It can also attempt to convert those parts into MIDI and teach you to play them through its KeyPrism trainer.
On paper, that's almost exactly the practice tool you need.
And after testing it, I discovered something rather unexpected: AudioPrism is very good at the bit that should be impossible, and surprisingly bad at the bits that shouldn’t be.
From "impossible" to "done" in 40 seconds
The desire to pull recordings apart certainly isn't new. Back in 2013, New Atlas reviewed Riffstation, software aimed at guitarists that could isolate parts of a finished recording and, crucially, let players solo or mute the instrument they were working on.
The separation technology was considerably more primitive, but the use case was already clear. Source separation has been developing for decades, but machine learning transformed what's possible. Deezer's Spleeter and Meta's Demucs helped make neural-network stem separation widely accessible, while Peter (Lord of the Rings) Jackson's team supplied a spectacular mainstream demonstration with its MAL technology. Developed for The Beatles: Get Back, it was subsequently used to separate instruments from the four-track masters for the 2022 remix of Revolver, and finally disentangle John Lennon's voice from piano on the old "Now And Then" demo.
Audio-to-MIDI followed a separate path. Ableton, for example, was converting audio melodies, harmony, and drums into MIDI with Live 9 more than a decade ago.
AudioPrism represents the convergence: unmix the recording, transcribe its constituent parts, then learn to play them.
Better still, and this is an element I personally find particularly appealing, the expensive computation happens on your own GPU. No uploading your music, no processing queue, and no metered cloud computation. That's not just good for privacy; it's wonderfully immediate.
Astonishing technology, baffling product
I gave AudioPrism two very different tests. It was very easy to operate; just drag and drop an audio file into it.
Fleetwood Mac's Dreams was the softball: an exceptionally clean studio recording with plenty of space around John McVie's bass. AudioPrism separated the entire song into four tracks on my RTX 5070 Ti in less time than it would take you to spell out loud "Fast Fourier Transform." Twelve-year-old me would have thought this witchcraft.
The bass was superbly isolated, and when I dropped all four stems into Audacity, recombined them, and A/B'd the result against the original recording, the transition was invisible. I was genuinely impressed.
Dreams is a fantastic track, but a lot less fun to play along to than with the genius of Tony Levin. So, for the bastard test, I used Red Rain from Peter Gabriel's Secret World Live: a dense live recording full of ambience, reverberation and overlapping instruments.
The extracted bass had considerably more artifacts and an unpleasant resonant/sibilant quality, but it was most definitely usable. More than good enough to practise with, a few minutes of cleanup would have made it pristine.
And that's when AudioPrism began driving me mad. Having produced these excellent tracks, where's the mixer?
As a bass player, my workflow couldn't be more obvious. Solo the bass while I learn the part, then mute it and let me play along with the rest of the band. Or similar for your chosen instrument.
Instead, I had to import them into Audacity and mute the bass myself. AudioPrism had performed the almost miraculous task of pulling apart a finished recording, then made me leave the application to accomplish the trivial bit.
As if to underline the point, while finishing this review I stumbled across StemKit, another local GPU-accelerated stem separator. And there, right in the application, was the thing I'd been asking AudioPrism for: a mixer. Separate level controls for every stem, with mute and solo buttons built straight in. No trip to Audacity required, and it’s free and open source!
It’s not the only example. In January, New Atlas covered JBL's US$599 BandBox Trio, a program that uses AI stem separation specifically so musicians can isolate or mute instruments and play along with the remainder of the track. In other words, the difficult technology serves the simple musical requirement, rather than becoming the product in itself.
I did not have a good time with the MIDI side. Feeding AudioPrism its own remarkably clean Dreams bass stem produced unpredictable multipart results rather than anything I found useful as a transcription. KeyPrism's falling-note playback also repeatedly stuttered and glitched, making meaningful assessment of its timing and accuracy difficult.
Then there's the interface. Oh my.
AudioPrism looks like it was designed by a hallucinating child who's just discovered Winamp visualizers. Glowing particles, mouse trails, pulsing graphics, and a permanently moving hyperspace backdrop compete for your attention; I couldn't find a way to turn them off. And, as the final insult, the hyperspace animation isn't even a clean loop!
I'm trying to concentrate. I'm trying to learn something. And it's doing its best to deny me.
The impossible bit is no longer the problem
AudioPrism also faces formidable competition: the aforementioned Stemkit. Ultimate Vocal Remover is also free and open source, supporting technologies including Demucs and MDX-Net. Moises approaches the problem much more like a musician: separated tracks sit in a mixer where they can simply be muted, soloed, or adjusted, alongside other practice tools. The catch is that it depends on cloud processing, and paid tiers expand what you can do. So, not local.
Even the local-processing advantage is no longer AudioPrism's alone. LALAL.AI's Lyra model performs stem separation on-device through its desktop app and VST plugin, although local processing requires its $19.99/month Pro subscription rather than a one-off purchase.
And perhaps that's the real story here. We've reached the point where pulling a convincing bass guitar out of a 50-year-old stereo recording in 40 seconds on a home PC isn't the problem anymore. What you build around that capability is.
That's what makes AudioPrism so frustrating. Its core separation technology can be spectacular, and doing all that work locally is exactly the kind of thing that floats my boat. But it repeatedly feels designed around what the technology can do rather than what a musician genuinely wants to do with it.
Give me AudioPrism's one-off cost, excellent local separation, a straightforward mixer with mute and solo buttons, reliable transcription, and an interface that doesn't resemble an 80s amusement arcade, and we'd be getting remarkably close to an ideal practice tool.
AudioPrism has already solved what looks like the impossible part. It's the ordinary bits that aren't there yet.
Gotta go. Pino Palladino's line from Sanctified has just finished cooking...
