XRB Audioxrb·audio

Acknowledgements.

XRB Audio, a trading name of XRealityBytes (United Kingdom) · Last updated 11 October 2026

None of this was made from nothing. The models in these products are fine-tuned from someone else's research; they learned from recordings made by hundreds of people who released that work for others to use; the software links code written by people who gave it away. This page names them, says what we used and under what licence, and says plainly where the record is incomplete.

The noise and music figures below come from the provenance record of the model that goes into the product — a per-clip licence audit of that corpus, bound by hash to the checkpoint rather than taken from a training script. The speech and singing figures are recounted from the corpus on disk and from each checkpoint's own training arguments, and the commit figures from the git history. Every licence named is the licence recorded for that specific clip, track or repository. The full per-clip lists are far too long for a web page. They are THIRD-PARTY-NOTICES.md and three companion files naming every contributor and every track, and they ship inside XRB IsoVox 0.2.0, the build made on 12 September 2026 after these corrections, and XRB IsoVox is not yet on sale. Ask us and we will send you the corrected files today.

What this is built on

DeepFilterNet

Both machine-learning models in XRB IsoVox — XRB_VX1_Speech and XRB_VX1_Music, renamed on 12 September 2026 from DFN3_Commercial40k and DFN3_MusicAug40k — are fine-tuned from DeepFilterNet3, in its low-latency configuration, by Hendrik Schröter, Alberto N. Escalante-B., Tobias Rosenkranz and Andreas Maier. That is the single largest technical debt this product owes anyone. The architecture, the deep-filtering method and the trained starting weights are theirs; what we added is a fine-tune on our own corpus and the work of getting it to run in a plugin.

The three papers the authors ask users to cite:

Repository: github.com/Rikorose/DeepFilterNet

The licence of those weights is an open question

The DeepFilterNet source code is generously licensed: the repository's README grants "All code in this repository" under either the MIT License or the Apache License 2.0, at the user's option, and every directory that ships as a package carries both licence files.

The published model weights are not covered by that sentence, and we cannot establish what does cover them. The directory holding the weights is not one of those packaged directories, and carries no licence file of its own; GitHub reports the repository's own licence metadata as NOASSERTION; and the downloaded archive contains only model files and a configuration file, with no licence or notice beside them. Three people have opened issues asking this exact question — numbers 697, 700 and 709, between 15 July and 1 September 2026 — and none has a reply from a maintainer. The last commit to the project's main branch was authored on 25 September 2024.

So: we have asked, and we do not have an answer. We are not going to describe that as a formality or imply a permission nobody has given. If the authors state a licence, we will comply with it — including by retraining from a different base, or withdrawing the models, if that is what it requires.

One further gap in our own records, for completeness: the project publishes the low-latency variant in ONNX form only, and our records do not document how the PyTorch copy we fine-tuned from was obtained. We would rather write that down than leave it looking settled.

The voices and the sounds

A noise-suppression model learns from recordings of real people singing and speaking, and from real sounds recorded by people who went out and recorded them. Everything below was released by its authors under CC0, CC BY or a public-domain dedication, except 49 recordings that are in the United States public domain by statute rather than by any author's grant, described below. Nothing licensed NonCommercial, ShareAlike, NoDerivatives or Sampling+ contributed to either XRB IsoVox model: material under those terms was removed before training. An automated licence check reads the corpus clip by clip and fails closed. It ran immediately before the music model was trained. It was written after the speech model was trained, so for that model the same clip-by-clip check is a later audit of the same noise corpus rather than a gate on the run, and we would rather say which is which.

Speech and singing — 74,588 files, 69.3 hours

Noise and music — 51,108 excerpts, 106.2 hours

Across FSD50K and Freesound, 4,769 clips carry a CC BY licence that obliges us to name their creator. All 4,769 are named, in a list of 1,229 distinct contributors — one of the companion files described above.

Where the credit stops, and why

Two things about that list are not complete, and we would rather say so than let it look finished.

A third correction belongs here too: until 12 September 2026 our notices credited the Cantoría Dataset to its institution and named neither creator, which CC BY 4.0 requires. Helena Cuesta and Emilia Gómez are named above and in the corrected notices — which, as said above, are not yet inside a build.

The software that ships

Open-source components distributed inside XRB Audio products, with the licence each is used under. A third-party notices file naming them ships inside every product bundle.

One more is credited here without being distributed: nlohmann/json — MIT — is required by the GPU Audio client headers, so it is reachable only from the GPU-accelerated build of XRB LinDyn, and that build is not distributed.

This website

Served to your browser or used to build the page you are reading: Next.js, React and React DOM (MIT), Three.js (MIT, which draws the wave on the home page), Stripe's and Resend's client libraries and Zod (MIT), NextAuth.js and its Prisma adapter (ISC), and Prisma's client and Postgres adapter (Apache-2.0). The typeface throughout is JetBrains Mono, by JetBrains, under the SIL Open Font License 1.1; this site self-hosts and therefore redistributes it, so its copyright notice and the full licence text are served alongside it at /fonts/JetBrainsMono-OFL.txt.

How this was built

XRB Audio is one person. The software was written with substantial AI assistance, and saying so is the same obligation as everything else on this page. Here is what the record shows, and you can check it against the commit history.

Across nine repositories between 4 July and 12 September 2026 there are 375 commits, counted immediately before the one that added this page. Of those, 329 — 88 per cent — carry a marker naming an AI assistant: 324 a Co-Authored-By trailer naming a Claude model (Opus 4.8, Fable 5 and Opus 5, in that order over the period), and 5 naming OpenAI Codex in the commit body. On every commit but five the git author is Greg Chapter; those five are GitHub-generated merge commits. No AI is the author of any commit in this project — only ever a named co-author.

The remaining 46 commits carry no marker. 42 of them fall in one repository across three weeks in July 2026, in the same style as the marked commits either side of them. The record cannot establish whether those were assisted or not, so we claim neither.

What the assistance did: wrote and revised most of the code, and — more than anything else in the log — checked it. The dominant pattern in the history is an implementation landed together with the measurement that verifies it, against a numerical gate, a parity check against the upstream reference, or a test. It also did the licence and corpus auditing this page rests on, clip by clip, including finding the 139-clip credit failure described above and correcting claims made earlier in this project's own work, including customer-facing text that overstated what the code did.

What it did not do: choose what to build, what to sell, or what to promise. The field testing on real equipment, the product decisions, the pricing, the legal positions and every release were the owner's. The rights in the software are held by XRealityBytes; the assistance holds none, and neither does any AI vendor. Where a question was a legal one, the log shows it referred back to the owner rather than answered.

We are not going to be coy about this, in either direction. A large share of this code was written with a machine's help, under a person's direction, and both halves of that sentence are load bearing.

Trademarks

Apple, the Apple logo, Mac, macOS, Logic Pro, MainStage and GarageBand are trademarks of Apple Inc., registered in the U.S. and other countries and regions. Google is a trademark of Google LLC. Other product and company names on this site are trademarks of their respective owners. XRB Audio is independent, and is not affiliated with, sponsored by or endorsed by Apple or Google.

Why this page exists

Ideas do not form in a vacuum. Everything here stands on research someone published, recordings someone made and shared, and code someone gave away — often with nothing asked in return beyond being named. That is a small thing to ask and an easy thing to get right, and the industry gets it wrong constantly, usually by saying nothing at all.

Recognising the originators of ideas is not a legal chore to be discharged in a file nobody opens. The sum of human knowledge should be a free resource available to everyone, and intellectual property that restricts the flow of ideas useful to people is a cost we all pay. The parts of this work we can give back, we will; the parts we sell, we sell while saying exactly what they are made of.

If you contributed to anything named on this page and we have you wrong, or missing, tell us and we will fix it — xrealitybytes@gmail.com. The same address gets you the full per-clip attribution files. Our licence terms are on the EULA and licensing pages.