Removing Noise with AI - RM Noise

[EDIT] I am periodically performing updates / maintenance on this system. If it isn't connecting, it should be back soon [/EDIT]


I have a new browser based client online for beta testing, and I am soliciting more feedback. Send your comments to: rmnoise@ournetplace.com (If you find an issue, consider documenting it with a screenshot.)

To run the new client:
a) In a browser, go to my new site here: https://rmnoise.com/
b) Click on "Web Client"
c) The first time, you will need to login. Afterwards, go right back to the above link and choose "Web Client" again.

Notes:

• You must already have a RM Noise username and password that was created before [edit] 1-4-2026 [/edit]
• Firefox and Chrome/Chromium have been reported to work on Windows 10/11 and Linux, but I do not know about Safari
• Right now, only 1 server is available, and it is in Texas
• The input device needs to support mono + 48kHz sampling
• You can see stats and logging on the browser's console (F12 on both Chrome and Firefox -> console)
• There is a lowpass filter mismatch between this client and the existing AI model. This *might* cause suboptimal filtering performance (I will need to retrain the AI filter, once I settle on the exact lowpass filter I want to use)
• Be aware of the "audio buffer" under options. Buffering is automatic, but if "late audio" is counting, raise "audio buffer" to compensate for the jitter (extra buffering comes at the possible expense of extra delay behind the radio's live audio)
• You can make recordings - they work, but they do not go to the online page yet
• I'm aware of a few cosmetic issues with the user interface
• You can click on the graph to replay audio from the click point (time-shift)

Works great on Windows 10, tested with Chrome, Edge, and Firefox. Trying Linux now.

73
Roland

I have also tested it on Ubuntu with Chromium and Firefox, all working ok. I am piping the Audio from a web-based SDR to RMN web client (in another browser tab) with the help of PulseAudio.

Roland

Short Demo RM Noise web-client on Ubuntu




Roland

Wow, the web client is fantastic!
I am sending audio from SDR Console.
Working great!

I stumbled across the web client by chance. I've been waiting for this for a long time
Thank you very much!

It runs perfectly with the following configuration:
- Fedora 43 with Null Sink Pulsaudio module.
- ExpertSDR3 as audio source.
- RM Noise as PWA on Chromium Version 143.0.7499.146.

Thank you for your input..!

I want to provide an update on the web client and how it relates to the user recordings.

The architecture of the web client is different than the windows client, and as a consequences, the neural network and training needs to be modified to match. I'm currently using all of my GPU resources to search for the best network and training parameters, but this search takes a long time. During this time, it is impossible for me to train the existing AI filter with new user recordings, but once the search has finished, I will go through the user recordings and train the production web AI filter.

It is hard to provide a timeline, but when the production web AI filter is [finally] available, I will retrofit each of the geographically diverse servers to also support the new web client.

To illustrate, each line represents an experiment that took between 1 and 2 days...

Works perfectly on Windows 11 and Brave Browser

The latest AI Phone filter is online (824e959). This filter matches the web client's audio characteristics.

Numerically, it is a strong performer.. Feedback is appreciated.

A new webclient version is now online (v0.421).

This version addresses glitches and pops in the filtered audio. This issue was present in all previous client versions!

Main changes:

• Fix for audio glitches and popping (happening with some, but not all, soundcards)
• Added the option to download audio recordings locally at 8kHz sampling
• Moved the stream processing to a webassembly module for better CPU utilization
• Changed dashboard to show input dBs (instead of output) and added SNR to the dashboard
• Cleaned up logging to browser's console
• Added additional telemetry from the client to the server
• Fixed a security risk at the server