[Feature request] Noise filtering feature mainly RNNoise

[ElrikPiro]

Context

For SDR (and other radio operating) users, noise seems to be a pretty common issue. Other solutions like the noise filtering plugin of SDR# allow users to filter different kind of audio tracks:

And I found it pretty useful, specially on bad signals setups.

Proposal

To add a noise filtering stage that allows users to select noise filtering profiles that allow users to digitally remove noise for a more clear audio experience.

[f4exb]

Alternatively WDSP receiver has a lot of digital noise reduction options based on the excellent wdsp library at the heart of OpenHPSDR: https://github.com/f4exb/sdrangel/tree/master/plugins/channelrx/wdsprx It has also a better chance to do a better job than just filtering audio because it has access to the full bandwidth of the original signal. It is particularly efficient with SSB signals but also covers AM and FM so pretty much all use cases where you need noise reduction.

I am not sure if a generic audio filtering feature will add significantly more to this. Of course it would be possible to pick up audio from a demodulator (like the demod analyzer does) filter it and render it to the audio output device. However is the code for all these noise reduction schemes publicly available?

[ElrikPiro]

I've been looking for it, as most SDR# plugins are available on github, the RTL-SDR blog quotes this RNNoise work as the one the AirSpy team used to implement that plugin, it seems to be using RNN to filter out noise from voices. I'll let you know if I find the actual source for the plugin.

Also, I'll check if the wsdspx plugin does the job, if the equalizer helps cleaning up Voice FM signals I would only suggest to have a combo with some preset configurations to ease up operation.

Thank you for your answer, sir!

[f4exb]

I have tried this RNN over several noisy samples (white noise, crowd noise...) and it seems quite efficient.

[ElrikPiro]

Are you considering actioning this into the sdrangel project then?

[f4exb]

I need to experiment a bit more to see if it is worth it. RNNoise is quite efficient at removing white noise or extracting a prominent voice from a crowd noise. I am not sure how this would apply in real life voice analog transmissions (digital voice is out of scope).

[ElrikPiro]

Thanks for taking the time for it

Please, keep me updated!

[f4exb]

It looks like RNNoise source code is mirrored on Github: https://github.com/xiph/rnnoise
Also this video seems pretty convincing: https://www.youtube.com/watch?v=pcpwFnj0Np4
Another one using SDR console NR3 which is RNNoise in the background: https://www.youtube.com/watch?v=YkV-L8vh7sQ obviously less convincing but it is a well known fact that the RNNoise model is good at removing background noise but as the signal of interest meets the noise it gets worse and worse.

[f4exb]

I conducted more tests and while it requires a reasonably high SNR to get something useful it completely cancels the noise and can reduce ear fatigue on loud enough transmissions with significant background noise so indeed it would be worth the effort to implement.

Basically there is nothing complicated: you feed it with 480 int16 samples from the audio channel (I think it is int16 already) and get 480 int16 denoised samples in return. This framing system is quite similar to other voice codecs.

The complication lies with the integration in SDRangel but I think it is nothing infeasible. The main points of concern are:

• No cmake build (autoconf, configure...)
• Windows support?
• Mac support?
• Possibility to download the archive of the set of NN factors and decompress it during the build. The corresponding .c file is quite large (~75MB) and cannot be reasonably fitted in a Github repository.

It does not seem complex at all to build and install the rnnoise library in a Linux system and compile SDRangel from source. The issue is with other forms of builds (workflows, other oses...)

[srcejon]

There are some open pull requests with Cmake and Windows support: xiph/rnnoise#88 - so looks doable.

[f4exb]

I will have to create my own fork. The cmake file in the PR does not take compile options into account so it produces unoptimized and unusable code. Moreover I would like it to manage the download and decompression of the model data file. The VLA replacement is a must though and I think not only for Windows

[f4exb]

Implemented