Donate Your Noise To Xiph/Mozilla's Deep-Learning Noise Suppression Project (xiph.org)

Mozilla-backed researchers are working on a real-time noise suppression algorithm using a neural network -- and they want your noise! Long-time Slashdot reader jmv writes: The Mozilla Research RRNoise project combines classic signal processing with deep learning, but it's small and fast. No expensive GPUs required -- it runs easily on a Raspberry Pi. The result is easier to tune and sounds better than traditional noise suppression systems (been there!). And you can help!
From the site: Click on this link to let us record one minute of noise from where you are... We're interested in noise from any environment where you might communicate using voice. That can be your office, your car, on the street, or anywhere you might use your phone or computer.
They claim it already sounds better than traditional noise suppression systems, and even though the code isn't optmized yet, "it already runs about 60x faster than real-time on an x86 CPU."

Comments

Sounds useful if

[Presence Eternal]

It says it can remove car noises, but can it remove the audience laughter from the Red Green Show? This is a problem someone needs to solve!

Re:Sounds useful if

[omnichad]

Have you tried duct tape?

Remember

[ArchieBunker]

Remember when Mozilla made a web browser? Pepperidge Farm remembers...

Re:Remember

[jmv]

(I'm the author of the article)
You may not be aware, but around 10 years ago, browsers started including audio technology. This now includes WebRTC which lets you do videoconferencing in the browser (without Flash). As surprising as it may sound, some people like doing VoIP/videoconference. And those who use WebRTC tend to prefer when their audio doesn't have too much noise. And that is why RNNoise is useful.

Re:Remember

[goose-incarnated]

No one uses WebRTC because it is in its infancy.

Every time I've tried to use it from behind a NAT I've had issues. It turned out to be the same issues that I had with SIP - not being able to get to the client due to a multi-layered NAT (two NAT layers or more). Most cell providers do double-NAT because for normal browsing and downloading it works just fine.

If you're doing SIP (or something like it, like WebRTC) you need an external proxy (STUN, TURN or ICE) server which all participants talk to, but which still won't work to get through the double-NAT.

WebRTC can't take off until everyone is on IPv6 and is not NATted. Or until they change the specification a little so that a connection-oriented (non-RTP) is specified. WebRTC peers can't talk to each other behind a NAT properly (only one party gets sound, for example), they have to talk to servers only. This needs to change or we need to move to IPv6.

Re:EditorDave

[CanEHdian]

Perhaps the poster is suffering from multiple personality disorder, so he can be Edit or Dave. When the Edit personality is dominant, everything is checked for spelling, style, punctuation and grammar. When Dave is dominant, less so.

What is your favorite noise?

[sheramil]

I suggest piping in a few tracks by SPK, in particular "Emanation Machine R.Gie 1916", the first track from their 1981 release "Information Overload Unit".

https://www.youtube.com/watch?v=G9b89PFYZ5g

When my wife first heard it, she said it was like having your head stuck inside a running vacuum cleaner. Follow it up with some Throbbing Gristle, perhaps.

Iffy

[fyngyrz]

I tried to donate noise; using a mac under 10.12.6. Mic is working fine. Safari asks if it can use the mic. The record button stays in for 60 seconds. The playback produces nothing.

I have great noise sources, and would not mind contributing.

Re:Data Set Publicly Available?

[fyngyrz]

The entire project is on github.

I found this by going to the link in TFS.

Re:Data Set Publicly Available?

[pthisis]

That link has only the source code. It does not include the training data set.

The submission link requires CC-0 attribution, which makes me hopeful that they plan to release the data freely. But I hunted all over the site and couldn't find either a link to the data or any comment about their plans for it going forward.

Re:x86 CPU?

[Desler]

Since when is ‘an’ equivalent to ‘all?’

Re:Data Set Publicly Available?

[jmv]

It's the first time we try this. We'll look at the quality of the data we get (yes, noise quality!) and if it's sufficiently good/useful, then we'll also make it available. It might take some time to sort out the useful samples from the ones that aren't since some already have noise suppression applied by the OS or browser.

Re:Data Set Publicly Available?

[jmv]

Now, training is a little trickier because I cannot share the data.

I cannot share the current data I'm using because it's copyrighted. Hence asking for people for help getting data that I can redistribute.

So weâ(TM)re supposed to just give jmv a bunch of data with no way to know how he is using it?

Yes, because I have such a track record for keeping things private.

Millions of Husbands...

[Captain Ramage]

Millions of husbands just submitted the sound of their wife's voice.