Show summary Hide summary
Musicians who want to isolate or mute parts of a finished recording now have powerful tools on their desktop — but not all of those tools feel designed for actual practice. AudioPrism, a Windows app that separates songs into individual tracks and converts parts to MIDI, shows both how far source separation has come and how uneven the user experience still is.
Why splitting a final mix matters
For players, the appeal is obvious: hearing a bass line by itself makes it faster to learn. Muting that bass so you can play along with the rest of the band is equally useful.
Historically, that was hard. Once a track was mixed to stereo, recovering the original instrument tracks was usually impractical. Advances in machine learning have changed that, turning a long-standing limitation into a routine task.
What AudioPrism promises and how it performs
AudioPrism is sold as a one-time purchase — listed at US$79.99 and currently available for $59.99 — and runs on a user’s PC, using the local GPU for processing. It separates recordings into stems for vocals, bass, drums and other instruments, attempts audio-to-MIDI transcription, and includes a practice mode called KeyPrism.
Operation is simple: drag an audio file onto the app and wait. In my tests, a clean studio recording — Fleetwood Mac’s “Dreams” — was split into four stems quickly on an RTX 5070 Ti. When I recombined those stems in Audacity and compared them to the original mix, the difference was practically imperceptible.
By contrast, a dense live song — Peter Gabriel’s “Red Rain” from Secret World Live — produced a bass extraction with more artifacts. It retained usable musical information, but required some cleanup to reach studio-level clarity.
Where the tool disappoints
After the impressive separation, the experience stalled. AudioPrism lacks an integrated mixer with mute and solo controls, so to practice I had to export stems and assemble them in another program. That interruption undermines the app’s main purpose for musicians: quick, hands-on practice.
The MIDI transcription and KeyPrism trainer also fell short. Feeding the isolated bass stem back into AudioPrism produced erratic multi-part transcriptions. Playback in the trainer stuttered and glitched at times, which made it difficult to judge rhythmic or melodic accuracy.
The user interface compounds the problem. The app’s animated background, visual effects and continuously moving graphics are distracting and showed no obvious way to disable them. For a tool aimed at focused practice, that felt like a mismatch between capability and design.
How it compares to alternatives
Audio separation and transcription are no longer niche experiments. Tools and models such as Deezer’s Spleeter, Meta’s Demucs, and Peter Jackson’s MAL pipeline have already demonstrated what neural separation can do. Ableton has offered audio-to-MIDI features for years, and open-source projects like Ultimate Vocal Remover provide accessible options built on these models.
Other products take a more musician-centered approach. StemKit, a local, GPU-accelerated project, includes a built-in mixer with mute and solo controls and is free and open source. Services such as Moises offer an integrated mixer and practice features but rely on cloud processing and paid tiers. LALAL.AI’s Lyra model can run on-device through its desktop app or VST, though local processing is gated behind a $19.99/month Pro subscription.
What matters now for practice tools
We’ve reached the point where extracting convincing instrument tracks from old stereo mixes is fast and practical on consumer hardware. That technical leap shifts the challenge: how to present those stems in a workflow musicians will actually use.
The ideal tool would combine AudioPrism’s local processing and one-off pricing with a simple mixer (mute/solo/level), dependable transcription, and an interface that stays out of the way. Until that balance exists, users will face trade-offs between accuracy, convenience, and privacy.
Pino Palladino’s bass line from “Sanctified” just wrapped up on my test playlist — time to see what I can learn from what remains of the mix.
