audio – How to get an updated ffmpeg arnndn noise filtering model?

In this question/answer, we see that an AI RNN model can be used in ffmpeg to remove background noise from audio.

Some model(s) trained in 2018 are provided by Gregor Richards here. In 2026, I didn’t find updates or variants of these models.

However, to build those, Gregor used RNNoise, a software provided by xiph.org and created by Jean-Marc Valin.

In 2018, Gregor made a few changes to the RNNoise software (he called it rnnoise-nu) and provided some trained models. The main/default model is cb and, in its ASCII version usable by ffmpeg, it is less than 300 kB in size.

Pre-trained updated Xiph models

Xiph.org also provides some pre-trained models to use with RNNoise: media.xiph.org/rnnoise/models/

Those are several MB in size, the last ones as of writing (I think, PyTorch weights and weights embedded in C source code):

• rnnoise10Ga_12.pth: 11.5 MB
• rnnoise10Gb_15.pth: 11.5 MB
• rnnoise_data.c: 78 MB
• rnnoise_data_little.c: 30.1 MB

As RNNoise is still maintained and updated (version 0.2 published in April 15, 2024), and the code and pre-trained models got further updates since then (in 2025),

• how can I use RNNoise (software and/or models) to update my ffmpeg arnndn model? (with what I hope could be a better quality model)
• how can I get different model sizes for arnndn? (say less that 1MB, and maybe other sizes if performance justifies)