For me noisetorch [0] (based on RNNoise [1][2]) works pretty good. I'm using RTX Voice while I'm working on Windows and noisetorch while working on manjaro.

RTX Voice seems to work a bit better, but noisetorch compares really well to it. I have a microphone which has static noise and is not really mechanical decoupled from my table. Both applications are good at cancelling that kind of noise.

Noisetorch is a bit more focused on suppressing noise while you're not talking, which is the use case of the posted article. RTX Voice is better at suppressing noise while talking, which you can hear in Nvidias demos as well.

[0] https://github.com/lawl/NoiseTorch

[1] https://github.com/werman/noise-suppression-for-voice/

[2] https://jmvalin.ca/demo/rnnoise/

It's worth noting that RNNoise uses way less resources than RTX Voice and doesn't require a GPU. There's a new paper by Jean-Marc Valin (RNNoise author) about a new and improved version PercepNet[0] that's also very light on resource usage, but it's currently lacking an open source implementation. I found someone working on that here[1]. So if anyone here has the time and ML chops to open source a PercepNet implementation, that would bring open source noise suppression quite a bit forward.

[0] https://arxiv.org/pdf/2008.04259.pdf

[1] https://github.com/jzi040941/PercepNet

If anyone knows how to configure pulseaudio to use both RNNoise and module-echo-cancel together, I'd very much appreciate some pointers. My results have been a very distorted robotic voice.

I love these solutions but none are as easy as "install an app and click a button to turn on." RTX Voice has that locked down. I'd use RNNoise instead if there was a standalone app that just "worked" on my mic input. But here I have to jump through hoops for it to work.

Can you elaborate on what hoops exactly you have to jump through? (genuinely curious, i wrote noisetorch and always looking to improve it)

Not the OP, but I'd say noisetorch already works pretty good for that use case. Especially with the recent additions of pre-selecting the last device and stuff like that, I really appreciate those minor QOL improvements.

If I compare it to how RTX voice works, it's _slightly_ more complicated to setup (currently no distribution through standard package managers/simply downloading a file and clicking it to start everything) and it doesn't autostart like RTX voice does on windows startup. But jugding from your issues [0] I understand the decision and I'll set it up manually.

It might be an idea to allow for autostart for users who don't care about the CPU impact because their CPUs are big enough or whatever.

Nevertheless, your solution is actually very close to what OP is describing: "install an app and click a button to turn on". So thanks for maintaining, I use it daily!

[0] https://github.com/lawl/NoiseTorch/issues/6

Compared to something like Krisp (since there's no RTXV on macOS) it was difficult because step one was to compile it.

Assuming the rest of the UX journey is similar (i.e. technically-oriented) compared to Krisp's menubar click-to-activate-mic UX it then becomes less approachable

On a tangent - nice work! That people are even comparing it to RTX Voice is impressive!

If you use GSuite Enterprise, you can use Meet's built-in noise cancellation which runs on TPUs in a DC.

It's basically magic. You can have someone vacuuming the room right next to you or cook dinner and your coworkers won't notice. Typing noise is just gone. It has improved our meetings so much.

https://krisp.ai works locally and is cheaper, but it noticeably degrades quality (tested with the Chrome extension only).

Can you describe this? Tried googling but couldn’t find due to TPUs being so ubiquitous.

As in what is a TPU? It's Tensor Processing Unit.

It's like a CPU/GPU, but specifically for doing the kind of math you come across with Machine Learning / Neural Network applications (mostly: lots and lots of dot products).

The noise cancellation :)

Krisp uses ML to remove background noises like dog barking, crispy chips and such.

Comes with a subscription model for unlimited use though.

https://krisp.ai/

I had people complain they couldn't hear me well when I used their (deprecated) iOS app. I kept having to turn off the ML noise-cancelling.

I had similar issues with their macOS app using a pretty basic headphone+microphone setup (EarPods). I ended up leaving it off and not re-installing in on my latest setup.

I've had a subscription from mid-2019, though I no longer use it. I find Zoom's basic noise suppression good enough for calls.

I believe Zoom licensed Krisp to handle noise-cancelling in their client.

Have you used the internal mic?

I found out that if the phone is too far away, the voice would flagged as ‚background chatting‘ and therefore muted.

Once I connected a headset those issues went away.

I was using standard corded iPhone earpods.

Any open source models available v

Here is a blog post that describes their approach https://towardsdatascience.com/real-time-noise-suppression-u...

Off-topic but I wonder if there's some creative department at Nvidia, whose job is to invent a compelling mass-market uses for their beefy GPUs, as crypto rush is fading out and the gaming market hasn't much potential to grow further.

I remember talking to a dude 10 years ago at an unconference whose grad project was to reimplement Database indexes to use GPUs. He was saying in some prototypes he was yielding orders of magnitude better performance, but it was still a very obscure idea.

This is the sort of stuff I'm sure they're already working on to find new customers.

Crypto rush is doing anything but fading. Eth mainnet has more hashing power on it now than it ever has. Nvidia is selling 30 series cards directly to mining firms, much to the chagrin of anyone who wants to buy one for themselves. https://etherscan.io/chart/hashrate

Only it's not a 'better way to go' because the thing doesn't work on the hardware this post is directly aimed at.

This seems like a fantastic use case for the ML hardware included on new-generation Macintoshes.