Repository navigation
Please state the license of the distributed model weights (rnnoise_data-*.tar.gz) #284
Description
Hi Jean-Marc, and thanks for RNNoise — we ship it inside a commercial Android app and it is the only denoiser that actually works on the low-end phones we target.
Our lawyer asked one question we could not answer from the repository: under what license are the pre-trained weights distributed?
What we found:
COPYING(BSD-3) covers the code. The weights come separately asrnnoise_data-<sha256>.tar.gzfrom media.xiph.org; the tarball itself carries no LICENSE, and neitherREADME,datasets.txtnor the release notes say what license the model is under. (The v0.2 note says the models are trained only on publicly available datasets, andmedia.xiph.org/rnnoise/README.txtputs the contributed noise data under CC0 — but that is the training data, not the resulting model.)- Distributions disagree on how to read this: Debian's
debian/copyrightand Gentoo's ebuild treatsrc/rnnoise_data.{c,h}as BSD-3 under the projectCOPYING; Fedora's spec tags the model asCC0-1.0. All three ship it, but none of them can point to a statement from you. - The specific model we ship is
rnnoise_data-0a8755f8e2d834eff6a54714ecc7d75f9932e845df35f8b59bc52a7cfe6e8b37.tar.gz(the currentmodel_versionon main).
Would you be willing to add one line — in README, in COPYING, or as a LICENSE file inside the model tarball — stating the license of the distributed weights (e.g. "the pre-trained models are released under the same BSD license as the code", or CC0, whichever you intend)? A reply in this issue would already help everyone who has to answer the same question.
A related point, if you want to address it: most of the clean-speech corpora in datasets.txt (the Google OpenSLR sets) are CC BY-SA 4.0. Our reading — and apparently yours, and Debian's and Fedora's — is that the trained model is not a derivative work of the training data and does not inherit share-alike. A sentence to that effect from you would close the second question lawyers tend to ask.
Thank you.
Activity
kei-kusanagi commentedon Sep 24, 2026
/// English ///
Could you also clarify the license of the official
rnnoise_data-0b50c45.tar.gzmodel?This is the model identified in the
model_versionfile at revision904a876dce1f9ab8860c0a5000ed151f9f6eef58(v0.2).File SHA-256:
4ac81c5c0884ec4bd5907026aaae16209b7b76cd9d7f71af582094a2f98f4b43.We would like to confirm whether these weights may be incorporated into and redistributed with a commercial desktop application, including when statically linked and distributed through app stores.
Are these weights covered by the same BSD 3-Clause license as the source code? If a different license applies, where can we find its terms and the notices that we must include?
The original question refers to a different model. We would appreciate explicit confirmation of whether the answer also applies to
0b50c45.Thank you.
/// Español ///
¿Podrían aclarar también la licencia del modelo oficial
rnnoise_data-0b50c45.tar.gz?Es el modelo indicado en
model_versionde la revisión904a876dce1f9ab8860c0a5000ed151f9f6eef58(v0.2).SHA-256 del archivo:
4ac81c5c0884ec4bd5907026aaae16209b7b76cd9d7f71af582094a2f98f4b43.Queremos confirmar si estos pesos pueden incorporarse y redistribuirse dentro de una aplicación comercial de escritorio, compilados estáticamente, incluyendo distribución mediante tiendas de aplicaciones.
¿Se aplica a estos pesos la misma licencia BSD de tres cláusulas del código? Si corresponde otra licencia, ¿dónde podemos consultar sus condiciones y los avisos que debemos incluir?
La consulta original menciona otro modelo; agradeceríamos confirmar expresamente si la respuesta también cubre
0b50c45. Gracias.ogra commentedon Oct 11, 2026
Related to the question above, but about a different point: the provenance of the noise and RIR data used to train the current weights (
rnnoise_data-0a8755f8e2d834eff6a54714ecc7d75f9932e845df35f8b59bc52a7cfe6e8b37).The README says the models are trained only on publicly available datasets. I could confirm the clean-speech datasets listed in
datasets.txt(OpenSLR, CC BY-SA 4.0 / CC BY 4.0) andrnnoise_contributions(CC0, per itsREADME.txt). I could not find source or license information for the following files underhttps://media.xiph.org/rnnoise/data/:background_noise_v2.sw,foreground_noise.sw/foreground_noise_v2.sw,contrib_noise.sw,synthetic_noise.swmeasured_rirs-v1,-v2,-v3Could you tell me:
Where these noise and RIR files come from, and under what licenses they can be used and redistributed?
Which versions of the speech, noise, and RIR data were used to train these weights?
Even a partial answer would help. Thank you.