Artificial Intelligence noise reduction

[Kurt V.]

Hi guys, 
please check my livestream of today where I demonstrate the newest AI noise reduction software. 

Before you listen to the livestream, it was not perfect, (some cracking and popping) because I was overdriving the input of the AI. But even with the overdrive, the results are absolutely amazing. 
Stream here: 

https://www.youtube.com/watch?v=kGB_BEc55pk

This is an absolute gamechanger.

We should ask the programmers of the Thetis software to add this extra noise reduction feature into their software. 
They should connect with Randy and see what can be done. 

The AI noise reduction virtually removes the ground floor noise 100%. It is in my personal opinion the ultimate filter for RX. 

The software can be downloaded here: 
 https://ournetplace.com/rm-noise/

You will need access and using a username and password by sending an email. A limited number of users will get access to the AI filtering server and service. 

I think a new Noise Reduction filter system is born, thanks to Artificial Intelligence. 

De ON7OFF. 

[Pez]

That's very interesting stuff. I have requested access and I am keen to see how it performs compared to Thetis NR2, and SDR Console's new NR5.
How did you find the extra delays it introduces? In your video, I could hear some low frequency "rumble" in the AI. Did you notice this artifact in person?

73 

[Kurt V.]

Not like in my livestream, but when the audio input is indeed at the correct level, you have zero ground floor noise. I know K1GMM uses for example a lot of workarounds with audio manipulation, noise gates etc. But here, the AI completely removes the hiss ground floor noise sound 100%. When a very low signal comes true, then you will have a readability of that station right away. 

Give me your experiences Pez. And if someone in the group has contact with the software developpers of Thetis, please ask them  to evaluate and implement this as an extra noise reduction feature in Thetis. 

It will make Thetis be the first using AI to filter RX. 

de ON7OFF

[Pez]

Ok, I have been testing RM Noise for a short while and the NR results are impressive so far. There is no significant artifacts like I heard in the livestream. The 400ms delay is a bit of an issue for me, but still the NR performance seems to be extremely effective.

For the CW guys, it has some very clever CW detection going on too! I would love to see some other voice filters, not just the 2.8kHz ones. 

My setup is Thetis HL2 Beta 5, sending RX audio via VB-Cable (virtual audio cable) to RM Noise. 

I'm not sure if direct integration into Thetis makes sense from a practical perspective? The results would be the same as running Thetis + RM Noise in the same way that I am now - that is, unless the AI processing can happen on the local machine, and not over the internet to the remote AI server (like it is now). But, I am not a developer so that assumption with a grain of salt! :) 

To clarify, I am not knocking this at all - it is very cool! Adding RM noise to *all* of your SDR software is as simple as using a VAC program (like VB cable), to send the audio to RM Noise as required. 

For those wanting to play with this:: 

Software: https://ournetplace.com/rm-noise/
Group: https://groups.io/g/RMNoise/

73

[Kurt V.]

Crasy, isn't it how it totally eliminates the ground floor noise. In my humble opinion, it is a game changer. Who can guide me to the email addresses who write the Thetis software. I want to connect these guys with Randy who makes the AI software. 

Maybe they can work together. Is there a group for the Thetis software? 

Any email addresses can be sent to my email you can find on QRZ.com

De ON7OFF

[Alan Hopper]

Hi,

I'm intrigued by this, sparksdr and I think sdrconsole have used code based on  RNNoise: Using Deep Learning for Noise Suppression - Mozilla Hacks - the Web developer blog for a while that seems to do something similar, have you any links with more detail of the technology? Is it the same , a development or something new?

73 Alan M0NNB

[ron.ni...@gmail.com]

I have the older RNNoise algorithm built into the TestFlight beta test builds of my iOS SDR app, which also supports using this NR when running an HL2. If anyone here wants to try it out, email me direct for a TestFlight invitation.  73, Ron, n6ywu

[DL1YCF]

Did I get this right, the NR works on audio samples, not on IQ samples?

If it works on audio samples, as has been discussed here in detail, 

it is an add-on that very easily can be added to any sort of program

Producing audio, including SDR apps.

If it works in „IQ space“ things are a little bit different, since you

can place it before or after AGC, and so on. Then it is not an extension to

Thetis itself but rather an extension to WDSP. 

There is an „extended“ version of WDSP that includes additional noise reduction,

namely RNNnoise and libspecbleach (search for VU3RDD and RNNnoise), and

piHPSDR includes support for such an „extended“ version of WDSP (compile with

the EXTENDED_NOISE_REDUCTION option). It is not a big deal to include in any

program that uses WDSP (piHPSDR, linHPSDR, Thetis, … )

Yours,

Christoph

--
You received this message because you are subscribed to the Google Groups "Hermes-Lite" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hermes-lite...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/hermes-lite/5223208f-e91f-4fc6-8b3e-c6ae23b46946n%40googlegroups.com.

[Kurt V.]

Is that software only available on Iphones, or do you also have a version for Android? 

I would love to test it Ron. 

Have a look at the recent livestream or the new video I have uploaded to my youtube channel (ON7OFF). 

Kind regards. 

[ron.ni...@gmail.com]

I currently have the RNNoise algorithm inserted in the audio output path (post demodulation) of beta test builds of my iOS SDR app.  RNNoise also has an VAD output, so can be used for AI/ML enabled squelch.

I don't do Android apps, as I neither have nor use those devices or their development tools. 

73, Ron, n6ywu

[Ramakrishnan Muthukrishnan]

If you like this running, I have a fork of pihpsdr/wdsp with rnnoise and libspecbleech integrated. I am use it with a radioberry based transceiver (a few of my friends are also using it on a daily basis).

The instructions to install can be deduced from this script (just take the compile instructions specific to wdsp and the two dependencies - rnnoise and libspecbleech) and pihpsdr fork. Christoph's version of pihpsdr should also work just fine as he wrote in this thread.

https://github.com/vu3rdd/radioberry-controller-pi-config/blob/master/pihpsdr_install.sh

In my experience, rnnoise works great when SNR is good. When snr is bad, rnnoise does not do a good job (ironic - that is when we really need to listen to the signal). libspecbleech does a better job in such cases. There are a few other noise reduction code out there which I have not tried.

If any of you like to try it out and run into problems, please feel free to reach me by email.

73

Ram VU2JXN (old callsign VU3RDD)

--

You received this message because you are subscribed to the Google Groups "Hermes-Lite" group.

To unsubscribe from this group and stop receiving emails from it, send an email to hermes-lite...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/hermes-lite/dcea03c5-4b71-4259-8604-b2be491c3637n%40googlegroups.com.

--

Ramakrishnan