Licences and credits

Last reviewed October 2026.

YourDevice runs on other people's work. Everything below either runs in your browser when you use a tool, or helped build the pages you are reading. We ship these libraries unmodified, as their authors published them. Most models had to be converted to run in a browser; where a model was changed on the way, its entry under The models says so.

The one that is copyleft: HEIC decoding

HEIC and HEIF support is libheif 1.19.8, which decodes the image itself using libde265 1.0.15. Both are licensed under the GNU Lesser General Public License, version 3 — a copyleft licence, stricter than everything else on this page, and it comes with real obligations. Here they are, met:

Notice. This site uses libheif and libde265, unmodified. The full licence text ships alongside the library itself, at /libheif/LICENSE.txt.
Source. libheif 1.19.8 — github.com/strukturag/libheif. libde265 1.0.15 — github.com/strukturag/libde265. The browser build we serve comes from libheif-js 1.19.8.
Your right to replace it. Section 4 of the LGPL says you must be able to swap in your own build of the library. We serve it as a standalone file at /libheif/libheif.wasm rather than baking it into our JavaScript, specifically so that you can. Replace that file and the converter keeps working.

The LGPL applies to libheif and libde265. It does not extend to the rest of this site.

Running the models

ONNX Runtime Web 1.27.0 (Microsoft) — MIT
Transformers.js 4.2.0 (Hugging Face) — Apache-2.0
WebLLM 0.2.84 (MLC AI) — Apache-2.0
kokoro-js 1.2.1 — Apache-2.0
Tesseract.js 7.0.0 — Apache-2.0. Its WebAssembly core also contains Tesseract, Leptonica, libjpeg, libpng, libtiff and zlib.
bergamot-translator v0.6.0 (Mozilla), the WebAssembly build of the Marian translation engine that Firefox uses for its own translations — MPL-2.0. We serve Firefox's own files unmodified, the WebAssembly and its JavaScript loader, at /models/mozilla/bergamot-translator/v0.6.0/, with the licence text beside them at LICENSE.txt. The source is github.com/mozilla/translations. The MPL applies file by file: to these files, not to the rest of this site.

Pictures and video

libaom, via @jsquash/avif 2.1.1 — Apache-2.0. Encodes the HDR AVIF output.
mediabunny 1.50.9 — MPL-2.0, file-level copyleft. Used unmodified.
gainmap-js 3.4.0 (MONOGRID) — MIT
gifenc — MIT · @monogrid, three.js — MIT

Documents

pdf.js 6.1.200 (Mozilla) — Apache-2.0. Bundles a JBIG2 decoder from PDFium (BSD), OpenJPEG (BSD-2-Clause), qcms (MIT) and a CC0 colour profile.
LibreOffice 24.8.8, compiled to WebAssembly — MPL-2.0, © the LibreOffice contributors and The Document Foundation. Word, PowerPoint and Excel to PDF run it in your browser. The build comes from the npm package @matbee/libreoffice-converter 2.7.2 (MPL-2.0); we serve its three engine files from /models/libreoffice/lo-24.8.8-yd1/. The WebAssembly binary and its data file are unmodified. We changed one string in the loader, soffice.js, so its thread pool starts 16 workers instead of 4:
```
-initMainThread(){var pthreadPoolSize=4;
+initMainThread(){var pthreadPoolSize=16;
```
The patch, the notices for the libraries inside the binary (PDFium, cairo, HarfBuzz, FreeType, ICU and others) and its bundled fonts, and the licence text are served beside the engine. LibreOffice's source is at git.libreoffice.org. “LibreOffice” is a trademark of The Document Foundation; YourDevice is not affiliated with it.
Noto Sans CJK (JP, KR, SC, TC), © 2014-2021 Adobe, under the SIL Open Font License 1.1. Fetched unmodified from /models/libreoffice/fonts-v1/ only when a document has Chinese, Japanese or Korean text.
pdf-lib 1.17.1 — MIT
@pdf-lib/fontkit 1.1.1 — MIT, using pako (MIT). Loaded only when a book needs a font the PDF standard fonts do not have.
EB Garamond 1.002, © 2017 The EB Garamond Project Authors, under the SIL Open Font License 1.1. Regular, bold and italic, shipped unmodified with the licence file beside them at /fonts/ebgaramond/. There is no Reserved Font Name. EPUB to PDF fetches it only for a book with Greek or Cyrillic in it, and subsets it in your browser — the OFL FAQ (1.12, and clause 5) expressly permits embedding a subset in a document, and doing the subsetting here rather than shipping a pre-cut font keeps what we serve an unmodified original.
qpdf 12.2.0 — Apache-2.0, © Jay Berkenbilt and Manfred Holger. Includes libjpeg-turbo 2.1.1 and zlib. (Its npm wrapper declares ISC; the binary itself says Apache, and the binary is what we ship.)
pdf-inspector (Firecrawl) — MIT, built on the Rust crate lopdf and around thirty others.
docx — MIT, using JSZip, which we use under its MIT option.
Hyphenation patterns, from the hyph-utf8 project — the same data TeX uses. EPUB to PDF loads the one matching the book's language and ships each file unmodified, licence header intact. American English © 1990, 2004, 2005 Gerard D. C. Kuiken, under a notice permitting copying and distribution in any medium without royalty provided the copyright notice is preserved; French — MIT; German — MIT; Spanish — MIT/X11; Italian — dual LPPL / MIT, used under MIT; Portuguese — BSD-3-Clause; Swedish — LPPL 1.2 or later.

Turning text into speech

Text to speech has to decide how every word is pronounced before it can say it. That step used to run eSpeak NG, which is GPL-3.0. It does not any more — the pronunciation now comes from two sources, neither of them copyleft:

The dictionary. The pronunciation lexicons from misaki — Apache-2.0, about 90,000 words for American English and 87,000 for British. We repack them; the words and their pronunciations are misaki's.
Everything the dictionary does not contain. Names, mostly. Those go through letter-to-sound rules from NRL Report 7948 (Elovitz et al., 1976, Automatic Translation of English Text to Phonetics) — a US Government work, in the public domain.
The text normalizer — the part that reads "3.75" as a decimal and "chunk.ts" as a filename — follows kokoro-js's, Apache-2.0 © the kokoro-js authors.

The six-language voices and voice cloning run Pocket TTS by Kyutai, licensed CC BY 4.0. Each stock voice carries the licence of the recording it was made from, listed under The models; two voices whose recordings are licensed for non-commercial use only are not offered. We changed the model: converted to ONNX and quantized to 8-bit with pocket-tts-onnx-export (MIT), so it runs in a browser. Kyutai's terms prohibit cloning a voice without explicit, lawful consent, which is why the tool asks for it.

Audio

The piano in Audio to MIDI is the Salamander Grand Piano V3 by Alexander Holm, licensed CC BY 3.0. The credit shown in that tool is a licence condition, not decoration.
RNNoise, via @shiguredo/rnnoise-wasm — the wrapper is Apache-2.0; the RNNoise model inside it is BSD 3-Clause, credited under The models.
The time-stretching and pitch-shifting in Speed & pitch is Signalsmith Stretch 1.3.2 — MIT, © Signalsmith Audio. Used unmodified, as the official Web Audio release, with its WebAssembly inlined in the library's own JavaScript.
The second copyleft one: MP3 encoding. Browsers cannot encode MP3 — Chrome's WebCodecs will encode Opus and refuses mp3, so the 320 kbps MP3 download is produced by lamejs 1.2.7, a JavaScript port of LAME, licensed LGPL-3.0 and used unmodified. That licence carries the same three obligations as libheif above, and they are met the same way. This is the notice; the source is linked; and your right to replace it is satisfied by architecture rather than by promise — the encoder is served as its own file at /lamejs/lamejs.js, with its licence text beside it at /lamejs/LICENSE.txt, and loaded from there at runtime rather than compiled into our JavaScript. Replace that file and the MP3 export keeps working. The LGPL applies to lamejs alone and does not extend to the rest of this site. Every other download this site offers is produced without it — the lossless WAV needs no encoder at all.

The models

Every model a tool downloads into your browser — 40 entries — with its licence, who holds it, and what was changed on the way. Each licence name links to its full text.

Speech recognition

Whisper tiny, base and small — MIT, © 2022 OpenAI. Converted to ONNX and quantised by onnx-community; we mirror their files unmodified. Source
Whisper small, Danish (whisper-small-danish) — MIT, © 2022 OpenAI; Danish fine-tune by thorhojhus. Converted to ONNX and quantised by us (the output layer given its own copy of the embedding weights so it quantises; identical values). Used when you transcribe Danish with the Quality model. The weights are MIT, but the model was trained on data with its own terms: CoRal v3 (a custom OpenRAIL-M licence from the Alexandra Institute and Alvenir), FT Speech (custom terms from the Danish Parliament), Common Voice and NST Danish (CC0) and FLEURS (CC BY 4.0), plus audio its author labelled with another model. CoRal’s licence carries use restrictions to models trained on it, so when you use this model you agree not to use it to imitate a person’s voice, or to infer anyone’s identity, gender, age, health or similar traits. Source
NB-Whisper small and base (Norwegian) — Apache-2.0, National Library of Norway (Nasjonalbiblioteket, NbAiLab). Converted to ONNX by us at a pinned revision. We re-quantised the 8-bit decoder after giving its output layer its own copy of the embedding weights (identical values), which halves its size. Used when you transcribe Norwegian. Trained by the National Library of Norway on about 66,000 hours of Norwegian speech: NRK subtitles, Storting (parliament) recordings, audiobooks and the NST database. Source
KB-Whisper small and base (Swedish) — Apache-2.0, KBLab, National Library of Sweden (Kungliga biblioteket). KBLab’s own ONNX files at a pinned revision. We re-quantised the 8-bit decoder after giving its output layer its own copy of the embedding weights (identical values), which halves its size. Used when you transcribe Swedish. Trained by KBLab on about 50,000 hours of Swedish speech: TV subtitles, Riksdag recordings (RixVox, ODC-BY), ISOF dialect archives and NST. Source
Distil-Whisper small.en — MIT, © 2023 The OpenAI Authors and The HuggingFace Inc. team. Converted to ONNX and quantised by onnx-community. Source
Silero VAD v5 — MIT, Silero Team. The ONNX export by onnx-community at a pinned revision; we mirror the file unmodified, with its licence next to it. Finds the pauses in a long recording, so it is cut there rather than in the middle of a sentence. Source
Nemotron 3 Diarization — OpenMDW-1.1, NVIDIA. Converted to ONNX and quantised to 8 bits by onnx-community; we mirror their files unmodified, with a copy of the licence next to them. Tells speakers apart, up to 8, for "Identify speakers". Source

Music and audio

Basic Pitch — Apache-2.0, © 2022 Spotify AB. Source
BS-PolarFormer — MIT, © 2024 Roman Solovyev (ZFTurbo), via Music-Source-Separation-Training. Converted to ONNX by bgkb (MIT); we mirror that file byte for byte. Source
RNNoise — BSD 3-Clause, © 2007–2017, 2024 Jean-Marc Valin; © 2023 Amazon. Built into @shiguredo/rnnoise-wasm (Apache-2.0). The BSD licence text and disclaimer are below. Source
DeepFilterNet 3 — MIT, © 2021 Hendrik Schröter. Converted to ONNX by soniqo (formerly aufklarer). DeepFilterNet is dual-licensed MIT or Apache-2.0; we use it under MIT. Source
resemble-enhance — MIT, © 2023 Resemble AI. Exported to ONNX by us, as five graphs from one checkpoint, and served from our own mirror. Nothing here came from a third-party conversion. Source

Text to speech

Kokoro-82M v1.0, including its voices — Apache-2.0, hexgrad. Converted to ONNX by onnx-community. Source
Pocket TTS — CC BY 4.0, Kyutai. Converted to ONNX and quantised to 8-bit by us, with pocket-tts-onnx-export (MIT), so it runs in a browser. Source
Pocket TTS voices from the CSTR VCTK Corpus — CC BY 4.0, Centre for Speech Technology Research (CSTR), University of Edinburgh. Each voice converted by us to the format our export uses. Source
Pocket TTS voice Alba — CC BY 4.0, Alba MacKenna. Converted by us to the format our export uses. Source
Pocket TTS voices from Kyutai voice donations — CC0, the donors, via Kyutai. Source
Pocket TTS voices from Voice-Zero (LibriVox) — CC0, LibriVox readers, via Kyutai. Source
Pocket TTS voice Estelle, from Kyutai's own recordings — CC0, Kyutai. Source
Pocket TTS voices from Common Voice — CC0, Mozilla Common Voice contributors. Common Voice clips are CC0; Kyutai publishes them in its CC BY 4.0 Pocket TTS repository. Source
Pocket TTS voices Jürgen and Rafael — CC BY 4.0, Kyutai. Licensed as part of the Pocket TTS repository. Kyutai does not say where these two recordings come from. Source

Pictures

ormbg — Apache-2.0, Maximilian Schirrmacher. Converted to ONNX by onnx-community. Source
Real-ESRGAN x4plus — BSD 3-Clause, © 2021 Xintao Wang. Exported to ONNX, one to one, by SceneWorks. The BSD licence text and disclaimer are below. Source
ATD 4xRealWebPhoto v3 — CC BY 4.0, Philip Hofmann. Exported to ONNX by us at a fixed 128 px input. The token sorting was rewritten as an equivalent counting sort so it runs on WebGPU, index types were narrowed to 32 bits, and one convolution was padded with zero channels. The weights are unchanged. The ATD network is by Zhang et al. (CVPR 2024). The upscaler’s Best detail mode. Source
Depth Anything V2 Small — Apache-2.0, the Depth Anything V2 authors. Converted to ONNX by onnx-community. Only the Small model is Apache-licensed; the larger ones are not, and we do not use them. Source
YuNet — MIT, © 2020 Shiqi Yu, from OpenCV Zoo. Source
Licence-plate detector (yolo-v9-t-640-license-plates-end2end) — MIT, © 2024 ankandrew (open-image-models). Source
Scratch detector from Bringing Old Photos Back to Life — MIT, © Microsoft. Exported to ONNX by us and served from our own mirror. Source
LaMa — Apache-2.0, © Samsung AI Center Moscow. Exported to ONNX by us and served from our own mirror. Source
DeOldify — MIT, Jason Antic. Exported to ONNX by us and served from our own mirror. Source
DDColor — Apache-2.0, © Alibaba (ModelScope). Exported to ONNX by us and served from our own mirror. Source
GMNet — MIT, © 2025 Yinuo Liao. Exported to ONNX by us, at fixed input sizes. Source

Documents

Tesseract recognition data (eng, osd) — Apache-2.0, Google and the Tesseract contributors. The tessdata_fast files, served unmodified. Source
LibreOffice 24.8.8 (WebAssembly build) — MPL-2.0, The LibreOffice contributors and The Document Foundation. One string in the Emscripten loader soffice.js changed: the thread pool starts 16 workers instead of 4. The patch and the notices are served beside the engine. Built by the @matbee/libreoffice-converter 2.7.2 package (MPL-2.0); the WebAssembly binary and data file are unmodified. Source
Noto Sans CJK (JP, KR, SC, TC) — SIL OFL 1.1, © 2014-2021 Adobe, with Reserved Font Name "Source". Fetched only for documents with Chinese, Japanese or Korean text, and served unmodified. Source
FFDetr — Apache-2.0, Joe Barrow (CommonForms). Exported to ONNX by us from FFDetr.pth (revision 56f4e42) with rfdetr 1.11.2 at 1024 px, and quantized to int8 at 768 px for browsers without a GPU. Served from our own mirror. The model card is tagged Apache-2.0 and the author has stated that licence; a LICENSE file has been requested upstream (jbarrow/commonforms#9). The FFDNet models from the same project carry no licence and are not used. Source

Language models and translation

Qwen2.5-1.5B-Instruct — Apache-2.0, © Alibaba Cloud (Qwen team). Run as MLC/WebLLM builds published by the MLC team, quantised to 4-bit. Source
Qwen3.5-4B — Apache-2.0, © Alibaba Cloud (Qwen team). Run as MLC/WebLLM builds published by the MLC team, quantised to 4-bit. Source
Firefox Translations models — MPL-2.0, Mozilla (the Firefox Translations project). Served unmodified from our own server (models gzipped for the download): the same files Firefox downloads from Mozilla. Source

The MIT License

Every model above marked MIT is licensed under this text, with its own copyright line as given in its entry.

Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.

The BSD 3-Clause License

Applies to RNNoise (© 2007–2017, 2024 Jean-Marc Valin; © 2023 Amazon) and Real-ESRGAN x4plus (© 2021 Xintao Wang). The licence asks that the copyright notice, the conditions and the disclaimer are reproduced with the binary, so here they are in full:

Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:

1. Redistributions of source code must retain the above copyright notice, this list of conditions and the following disclaimer.

2. Redistributions in binary form must reproduce the above copyright notice, this list of conditions and the following disclaimer in the documentation and/or other materials provided with the distribution.

3. Neither the name of the copyright holder nor the names of its contributors may be used to endorse or promote products derived from this software without specific prior written permission.

THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

The site itself

Astro, React and React DOM — MIT. DOMPurify is offered under MPL-2.0 or Apache-2.0, and we use it under Apache-2.0. marked — MIT. KaTeX 0.18.4 — MIT, which typesets the maths in Private chat, fonts included. Plus Jakarta Sans by Tokotype — SIL Open Font License 1.1.

Our own code

Spotted something wrong here? A mistaken licence claim is a bug and we would rather fix it — get in touch.