Table of Contents Show
For those who don’t know the tool: iZotope Photon is a standalone, GPU-accelerated application and ARA2 plug-in that separates elements of existing audio and repairs dialogue using locally running machine-learning models. For picture post, it offers a way to rescue compromised production sound or rebalance a finished mix without asking the sound department to reconstruct history.
There are two kinds of people in post-production: those who understand audio, and those of us who confidently click the little speaker icon and hope the problem becomes someone else’s. iZotope Photon is for us in the second group. It brings stem separation and noise reduction into one interface, with twelve custom-trained models and controls that do not require a degree in acoustics or a close personal relationship with a compressor.
Photon is the first joint engineering effort from iZotope and Boris FX, succeeding the product previously known as CrumplePop SoundApp. The Boris FX team has been shifting audio functionality between products and packages as the two portfolios come together. That does not mean the tools have disappeared: the capabilities remain and are continuing to develop, now with Photon as the new home for the SoundApp lineage and additional machine-learning models. For existing users, the key is to check which current package and model covers their workflow, and not assume a name change means their tools have gone missing.
Where it fits into picture post
The obvious use case is not making a karaoke version of a song, although Photon can do that too. It is the documentary interview recorded in an excessively reflective room, the presenter competing with Traffic, people speaking over each other on a remote call, or a locked promotional video whose Music is louder than its message. When sound elements are baked together, conventional volume adjustment changes them all at once. Source separation tries to recover some control over individual components.
The Cinema 4 Stems model is central here. It takes an audio or video file and separates its soundtrack into voice, Music, effects and other content. A picture editor could therefore reduce the Music beneath a last-minute line of voiceover, attenuate effects or extract dialogue for further treatment, then export the result. Separation artefacts and leakage are possible, so critical listening still matters. Unfortunately, that part cannot yet be delegated to a button.
The everyday lifesaver: removing Music from a finished video
The most immediate editorial use may be removing a music bed from a finished video while keeping the voice and original effects. Imagine the Client’s locked promo: the music clearance expires, and someone asks for a version with different Music by lunch. Naturally, the original edit project and separate audio tracks are on an archive drive no one can find. This is exactly the kind of request that makes us consider retraining as a camera tripod.
For a copyright-driven replacement, the sequence is simple: import the source MP4, MOV, MKV, MXF or MPG/MPEG; install or select Cinema 4 Stems; audition the separated voice and effects; mute or reduce Music; export individual WAV stems, a revised stereo mix or a same-as-source video file. The demonstration says the picture stream is retained without re-encoding. That is especially valuable when the only available asset is already approved visually and a last-minute audio change must not alter the picture.
There are two important caveats. First, this is model-based separation, and the amount of music bleed or audible artefacts will depend on the source. Every dialogue line, transition and exposed effect needs a listening pass. Second, removing an unlicensed music bed is an editorial remedy, not automatic legal clearance for everything else in the source. The final deliverable’s rights position still needs to be checked.
The same workflow also helps make a social cut from an existing trailer, pull a usable line of dialogue for a new promo, recover sound effects for another edit or duck the score under an important sentence. And because the models run locally on the GPU, the demo doesn’t require sending the video to a cloud service. Those of us who failed audio class can at least retain custody of the files.
Dialogue rescue: from noisy streets to glass-walled meeting rooms
Photon’s Denoise family is organised by source problem rather than by a traditional chain of audio processors. Voice Restore combines controls for Strength, Level, Mic Tone, Remove Room and Remove Noise. Its demonstrations include a piece to camera on a busy street and a corporate interview with room reflections and thin microphone tone. Level control can even out a voice, Remove Room targets reverberation, and Mic Tone adjusts the perceived fullness of speech. These controls aren’t interchangeable: an echo problem isn’t the same as steady noise, no matter how enthusiastically we may have treated both with the volume knob in the past.
The other noise models are Indoor, Indoor Noisy, Outdoor, Outdoor Traffic and Outdoor Wind. That makes six denoise variants when you include Voice Restore. This selection suits news, documentaries, branded content, and location-heavy productions, where the editor may inherit a wide range of recording environments. A sensible workflow is to select the model for the actual problem, audition a short representative passage, adjust conservatively and check consonants, breaths and background transitions before committing to an export.
Separating people who insist on talking simultaneously
Voice Separator can split up to six speakers from a single file into separate stems. The demonstration uses a remote conversation in which latency or headphone trouble causes one participant to start before the other has finished. The operator specifies the speaker count; Photon then provides a separate row for each detected voice, with level, solo and mute controls, plus an Other row for remaining background content.
For interview editing, this is potentially more valuable than any music feature. Isolating two overlapping speakers can make an otherwise unusable exchange easier to cut and process.
The twelve models, without the fog
Demix: six ways to disentangle a recording
- Cinema 4 Stems: dialogue or voice, Music, sound effects and other material from audio or video.
- Voice Separator: up to six individual speakers, with an additional Other component.
- Voice 2 Stems: voice versus remaining musical material, useful for extracting speech or vocals.
- Music 4 Stems: vocals, bass, drums and other instruments, optimised as a quicker option in the supplied tutorial.
- Music 4 Stems Ultra: the higher-quality four-way music separation model.
- Drums 5 Stems: kick, snare, hi-hat, toms, and cymbals, primarily for music editing and performance work.
Denoise: six source-oriented cleanup models
- Voice Restore: speech enhancement with level, microphone tone, room-removal and noise-removal controls.
- Indoor: room-focused dialogue cleanup.
- Indoor Noisy: indoor speech with additional background-noise problems.
- Outdoor: general exterior recordings.
- Outdoor Traffic: traffic and traffic contamination.
- Outdoor Wind: wind-heavy exterior material.
The distinction between Music 4 Stems and Music 4 Stems Ultra is that Ultra targets cleaner separation, while the regular model is the faster alternative.
GPU processing and the practical workflow
The machine-learning models run locally on the workstation GPU. So processing doesn’t require uploading audio to a remote service or using online processing credits. This is especially appealing for confidential interviews, unreleased productions, and material that cannot be casually sent to a cloud service. Local processing isn’t the same as a blanket security certification, but it removes a major data-transfer step.
Photon begins processing when media is imported, and preview can start at any point in a clip. For a long interview, the operator can jump to the worst section and hear a result without waiting for the entire recording to be prepared. The application also lets you load and switch between multiple models, and the Model Manager downloads new or updated models separately from application reinstallations.
The standalone application opens WAV, MP3, FLAC, AIFF/AIF, AAC, M4A and video formats including MP4, MKV, MOV, MXF and MPG. It reads the audio from a video file and can export processed audio back into a same-as-source video while retaining the original video stream. Audio export options include WAV, FLAC and MP3. For picture editing, that means a direct route back to a video asset or an audio-only interchange file, depending on the facility’s workflow.
An ARA2 plug-in version is supplied for compatible digital audio workstations. The hosts include Pro Tools, REAPER and Cubase, with Sequoia and Samplitude additionally specified for Windows. It would be funny if those two Boris FX Tools were incompatible. But they aren’t.
ARA gives the plug-in timeline access in those hosts and supports adjustments without relying exclusively on standalone round-tripping. The material does not establish native Photon plug-in operation inside DaVinci Resolve, Premiere Pro or Media Composer. For those editors, use the standalone import and export route.
Hardware and platform requirements
On macOS, Photon requires macOS 14 or later and Apple silicon (M1 or newer). Intel Macs are not supported. The supplied specifications recommend 16 GB unified memory and list 8 GB as the minimum. Processing uses Core ML.
On Windows, requirements are Windows 10 or later (64-bit); an Intel Core i5 (fifth generation or later); and a GPU with at least 4 GB of VRAM. System RAM is 8 GB minimum and 16 GB recommended. Both platforms support ARA2 in compatible hosts. GPU acceleration is central to the design, so check hardware suitability before treating the application as an emergency repair station on an older editing machine.
Pricing, upgrades and availability
The supplied launch pricing lists Photon at US$349 MSRP, with a US$259 introductory offer through 7 December 2026. Crossgrade pricing depends on the products already licensed: any paid iZotope product, US$269 regular or US$179 promotional; RX Elements, US$199 or US$129; RX Standard, US$149 or US$74; and RX Advanced, US$99 or US$49. Those offers include the qualifying RX bundles.
Subscriptions are at US$19 per month or US$119 per year. The introductory offers are US$12 for the first month, followed by US$19 monthly, or US$89 for the first year, followed by US$119 annually. Exact account-specific prices appear after you log in to iZotope.
Existing customers with an active CrumplePop subscription, Boris FX Suite subscription or CrumplePop upgrade and support plan are entitled to Photon at no additional charge. They can retrieve its installer through the Boris FX Hub.
| Tool | iZotope Photon |
| Developer | iZotope / Boris FX |
| Product lineage | CrumplePop SoundApp |
| Function | AI stem separation, dialogue enhancement and noise reduction |
| Models | 12, comprising six Demix and six Denoise models |
| Formats | Standalone application; ARA2 plug-in for compatible DAWs |
| Processing | Local GPU acceleration, without audio uploads in the described workflow |
| Inputs | WAV, MP3, FLAC, AIFF/AIF, AAC, M4A, MP4, MKV, MOV, MXF, MPG |
| Outputs | WAV, FLAC, MP3 or same-as-source video with original video stream preserved |
| macOS | 14+, Apple silicon M1+, 8 GB unified memory minimum, 16 GB recommended |
| Windows | 10+ 64-bit, Core i5 fifth generation+, 8 GB RAM minimum, 16 GB recommended, GPU with 4 GB VRAM minimum |
| Regular price | US$349 MSRP; US$259 introductory offer through December 7, 2026 |
| Subscriptions | US$19 monthly / US$119 annually, with temporary first-term discounts |
| Eligible CrumplePop / Boris FX users | No-charge Photon entitlement for specified active subscriptions or upgrade and support plans |
| Product | iZotope Photon |