iZotope has released Photon, its first plugin since Boris FX acquired the company in July, and it is a dedicated stem splitter. According to MusicRadar's write-up, Photon runs as a standalone app and as an ARA plugin, carries 12 custom-trained machine learning models that run locally on your GPU, and exports stems as WAV, FLAC or MP3. The models include a four-way music split (Voice, Bass, Drums, Other), a drum model that separates a kit into Kick, Snare, Hi-hat, Toms and Cymbals, a voice model that pulls an acapella, and several post-production models for speech cleanup and separating dialogue, music and effects from video. The models were built with the Boris FX team. New models download through a Model Manager window without reinstalling.
If you have used a stem splitter before, the shape is familiar. What is worth reading closely is the drum model, because that is the part that touches anyone trying to get a score out of a recording.
Why a stem splitter matters when the goal is notation
Most people who want notation out of a recording start with a mix: a full song, a live board tape, a rehearsal phone recording. A transcription model asked to read that mix has to pick out the bass line through a wall of guitars, or find the kick under a compressed master. The cleaner the audio you feed it, the fewer wrong notes you fix afterward. That is the whole argument for separating first, and it is why we built a stem splitter into the same pipeline as the transcription.
Photon's four-way split is the standard shape. The drum model is the interesting one. A single "Drums" stem still contains kick, snare, hats, toms and cymbals stacked on top of each other. A drum transcription model reading that has to decide whether a low thump is a kick or a floor tom, and whether a wash of noise is a hi-hat or a crash. Split those into five stems and the question gets easier: the kick stem has kicks in it, the hi-hat stem has hats.
For a drummer learning a part, or a teacher writing out a groove for a student, that is the difference between a draft with a dozen misread hits and a draft with two or three.
What changes in practice
| What Photon offers | What it means for a score |
|---|---|
| Five-way drum split (kick, snare, hi-hat, toms, cymbals) | Cleaner input for drum transcription; fewer kick/tom confusions to fix |
| Four-way music split (voice, bass, drums, other) | Standard prep before bass, guitar or piano transcription |
| ARA plugin in a compatible DAW | Separate on the timeline without leaving the session |
| Local GPU processing | Your audio stays on your machine |
| WAV, FLAC, MP3 export | Stems drop straight into a transcription tool or a DAW |
| Model Manager | New models arrive without a reinstall |
The ARA point is worth spelling out. If Photon runs inside your DAW, you can separate a track on the timeline, mute the stems you do not need, and work from there. For a producer building a remix or a session player learning a part, that removes an export-and-reimport step.
What it does not change: Photon outputs audio. It does not produce notation, MIDI or a score. You still need a transcription step after it, whether that is a person with a pencil or a tool like ours. The separation is the prep, not the deliverable.
Our view
The honest trade-off with any stem splitter is that separation is not free of artifacts. Models leave bleed, smearing and the occasional ghost note, and a transcription model downstream will read those artifacts as real notes if you are not careful. A five-way drum split can make drum transcription easier, but it can also hand you a hi-hat stem with snare bleed in it. The fix is the same as always: listen to the stem before you trust the transcription, and edit the result.
The Boris FX ownership angle matters less to a working musician than the model lineup, but it is worth noting. iZotope's RX and Ozone already had stem separation; Photon is the first dedicated tool, and it was built with the new owner's team. That suggests the separation models are a strategic product line rather than a checkbox feature, which is usually good news for how often they get updated. The Model Manager window supports that read: models are treated as things that ship and improve, not things frozen into a release.
What we cannot tell you from the announcement is how Photon's drum split compares to other splitters on real recordings. We have not run that test, and neither should you take a press release's word for it. If you are choosing between tools, compare on verifiable facts: which instruments each one separates, what formats it exports, whether it works as a plugin or standalone, and how it is priced. Then run both on the same busy mix and listen.
Who should care: drummers and drum teachers working out parts from recordings, transcribers and arrangers who currently hand-split a mix before writing it out, and producers who need stems for a remix and notation for the session. If you only ever transcribe a single solo instrument recorded cleanly, a stem splitter adds a step for little gain.
One thing the announcement does not address, and neither will any tool: separating a commercial track does not give you the rights to distribute what you make from it. A transcription of a copyrighted song is a derivative work. If you plan to publish, sell or perform the result, check with the rights holder or your local licensing body. The rules differ by country.
What to do next
If you already have a splitter you trust, the question is whether Photon's drum model or its ARA workflow solves a problem you actually have. Test it on a recording you know well, ideally a busy one with a real kit, and listen to the hi-hat and tom stems for bleed before you judge the split.
If the end goal is a score, the workflow is the same whichever splitter you use: separate, then transcribe, then edit. You can transcribe a song, run the audio to MIDI converter when you want the part in a DAW rather than on a page, and browse shared transcriptions to see what other people's splits and transcriptions look like. Plans and what they include are on the pricing page; the shape of the offer is a free tier to try the pipeline and paid tiers for longer or more frequent work.
Automatic transcription is a first draft. Photon can make that draft better by giving the model cleaner audio. It cannot make it final. That part is still yours.
