You can now turn on Nvidia's excellent noise cancellation with any GeForce GPU | Hacker News Reader

You can now turn on Nvidia's excellent noise cancellation with any GeForce GPU

pcgamer.com

109 pointswoliveirajr65 comments

Is it just me, or is a proper headset with a quality directional boom microphone a superior solution to majority of software noise cancelling?

I hate to sound like a luddite, but I'm on conference calls all day every day; zoom, webex, MS teams, skype, you name it; and the more laptop-microphones there are, the mushier and laggier everything and everybody sounds due to all the noise cancelling (and crappy microphones / noise environments to start with). :-/

I have three headsets. One is a top recommended Bluetooth mono earpiece that cost ~$100 or so, one is a fancy ‘office’ usb headset from a big name that cost (my employer at least) $200+ (But has the mics built in to the ear cups), and the third is a $10 old school analog headset. It looks dorky, and it isn’t the most comfortable thing ever, but I get the best video conferencing results with the $10 headset because it has a boom mic that brings the mic close to my mouth.

Oh absolutely; Boom makes ALL the difference. If it doesn't have a boom, it's a headset with microphone as afterthought which by definition is setup to get all the noises from everywhere. The big headphones with microphone in earpiece are particularly horrible - I've used it couple of times out of desperation and people could hear my baby or wife two rooms over more clearly than myself :-/

The Logitech H151 is $20CAD when on sale, and though it's a bit "crisp/sharp", sounds way better than the Beats And Boses and Apples and Samsungs and Sony non-boom microphones people try to repurpose :-(

Several years back a previous employer bought us all Jabra 65 headsets. Can be connected via bluetooth or a little USB dongle (which is probably just bluetooth).

Best noise cancelling mic I've ever used. Someone can be talking right next to me and nothing will go over the mic except my voice talking directly into the boom.

Down side is you have to have the boom directly in front of your mouth; not slightly below or above.

After iterating quite a bit, I settled on using open back HD650’s and attaching a ‘mod mic’. The wire management is actually easy. This feeds into a DAC / amp combo from Schitt, mines the Fulla Schitt gen 3. Never had a better audio setup, for music, conferences, and games. Swap in a comparable closed back pair of headphones if working in a shared space.

Considering a custom rolled tube amp too just for the fun of it....

Proper headset is the first step. Just by everybody having good headsets and small lag -> everybody wins.

But in some situations having the extra help is what makes the difference. For example: today my room is noisy because some neighbour is using a drilling machine. With those softwares, I'm helping everybody else in my meetings, even if my life is miserable right now.

Proper headset and/or proper microphone training. A lot of people just don't know, or care.

You know this, and I know this, but 99% of people who use video calling software will never know this, so the software vendors are playing the hand they are dealt and improving the parts that are under their control.

Laptop microphones are obviously problematic, but boom mics have their own issues. None of my coworkers know how to position the microphone properly, so whenever one of them talks it's riddled with plosive whooshes of air whenever they make a "P" or "B" sound.

I think videoconferencing software should have a mandatory "preview room". Well not mandatory, but like

  [preview microphone/camera] [join meeting]

Sort of like a staging area to check if your laptop actually works with the projector before you are in front of people.

You mean like meet has?

I was thinking teams or webex or the software large organizations use. (I'd love it if people talking to the group using the laptop microphone in their tiled kitchen could hear how they sound)

Depends, I have a very good boom mic plugged into a solid interface that has essentially no background hiss or noise.

The software noise cancelling like Krisp in Discord or this Nvidia software is really nice because it stops outside noises like dogs barking, keyboard keys, etc from coming through.

With Krisp enabled in discord I can eat chips with voice activation enabled and no one can tell, and no one can hear my dog bark occasionally even if I'm talking while it happens.

And don’t forget the bane of working from home in the suburbs: leaf blowers.

It seems at least once per day there is an active leaf blower in my neighborhood vicinity. Same goes for my colleagues.

Noise cancellation is great for, well, cancelling their noise.

The worst, especially where I live in the SW. Most of the trees are evergreen so there was no relief during the winter. My office window overlooks five neighbor yards and I have each one's lawn care schedule on my calendar.

Don't apologize. Quality hardware is an underrated boost to basically everything in life, as long as you can afford to buy it.

Continuing on this tangent, my office bought headsets for all employees so we could remote work. Bought dozens of Sony WH-1000X M3, turns out the microphone is ridiculously bad on these.

I’ve found if you have the headset in noise cancelling mode it messes with the mic and makes it worse. They may have patched it but was bad enough I stopped using them too.

I have the XM2 and the main issue is using the microphone on a laptop drops it to the “hands free” bluetooth profile where both the mic and audio out use a super low bandwidth codec that sounds awful.

When I disable the mic the audio quality goes way up as it switches back to a higher bandwidth codec. It’s super annoying.

I’m not sure how the apple airpods get around this, I think something to do with their custom chip.

AirPods also drop quality when the mic is in use!

I don't get it. They are wireless, noise-canceling headphones, not communication headsets. They are super expensive and I'm sure excellent for music/listening, but have no boom and are in no way meant to be used for serious/frequent calls. Heck, their tagline is "only music; nothing else". If THAT's not a hint, I don't know what is...

It's depressing to think that a $20 headset may have been a better investment than the $300 headphones :<

Being good as headphones does not necessarily make a headset bad as a microphone.

I have a set of Bluetooth Plantronics Backbeat PRO. There is no boom mic, but everyone I have talked to on cell phone with them has said they sound good. Even when I am out for a walk they filter out all the vehicle noise and wind.

Hmm; that's fascinating - I am skeptical only because I happen (no kidding:) to own it, and they are the example I used in another comment - people can hear me typing, my wife in another room, etc. When I'm outside, they multiplied the wind as it gets caught between the thick headband and my hair and then picked up by microphone :-/

(I also tend to speak too loud with them, as they block outside noise effectively, but that's a "user error" thing :)

I enjoy them as headphones - convenient, simple to pair, comfy - but they are specifically the reason why I pursued a boom headset and why I'll never go back :)

Without testing I cannot speak to our differing experience; I DO notice that some friends/colleagues of mine have a very very low threshold of what "sounds good" - if they can even remotely understand me, it's "Fine". They are also the ones who don't mind 30 people joining with laptop mic/speakers and the muffled/delayed/garbled sound that results... Others provide much more critical feedback (calling my 75 year old father is a great way to test a new headset :P )

Or, you may have found a button I haven't! :=>

Yes to everything you said. My employer gave all employees a webcam with a half decent microphone built it, made a huge difference.

Related, and it could just be me, it seems that those who call in from their phone have some of the worst sound quality. Could just be a small sample size though.

I have a fancy mic that I keep close to my face, but you still pick up some noise in an untreated room, so some sort of noise gate is important. (I personally use Krisp, which works pretty well. You hate to use a subscription service for something that could be done in hardware, but it's a lot of months of $5/month before you break even, and software doesn't take up valuable desk space with knobs that you never turn.)

To me, built-in echo cancellation tuned for people that do video calls with speakers / laptop mic is the worst thing. It tends to mute your audio for the first few seconds of your "turn". That means if you are just saying "yeah" and nodding, an important component of how humans interact, you disrupt the meeting for 5 seconds instead of adding some social smoothness. Use headphones and turn that off, and everything is a lot better. (If you use Zoom with a good audio setup, you have to turn on "original audio" and then turn off echo cancelation, and make sure you aren't echoing.)

But yeah, with any sort of sampling task, you want to increase the signal to noise ratio. The best way to do that with a microphone is to have it close to your mouth. A $4000 microphone on your desk is going to sound worse than a $70 microphone close to your mouth. That's just physics. Building a microphone into your keyboard is basically the worst thing you can do.

I find dynamic mics the best. Their ability to pick up noise drops rapids as you move away from the mic.

Yes, but most people use laptops and not desktops.

So there needs to be a good solution for them without a boom mic. But also, a boom mic with quality noise canceling (a lot sucks right now, kinda like the backgrounds) will still be useful.

The RTX limitation was always a sham, and editing some config files allowed the feature to work on any CUDA-capable GPU from the start.

Removing this arbitrary limitation is a good thing, but they should have done this earlier.

If there was strong competition, none of that circus would be possible.

Arguably nvidia has been innovating while on top to prevent the competition that does exist to catch up.

They have such a dominant market share because they make more compute for less dollar. When the day comes that isn’t the case, they’ve already made well established hardware features and adopted APIs people want: raytracing and machine learning.

Beyond that they're doing some truly spectacular stuff with their hardware. DLSS is amazing quality-wise and lets you run games at acceptable quality with much weaker hardware than it would take to run at native resolution, and because it's primarily implemented using the dedicated ML hardware, for most games it doesn't eat into valuable rasterization/compute resources. AMD simply does not have the hardware to do this right now and they also don't have the years of research and development that went into modern DLSS (like the generally crappy 1.x versions of DLSS)

While true quality fanatics are going to want to run things at native resolution, the average person will be quite happy with DLSS 4k instead of native 4k, and they'll be able to get much higher framerates in that mode than on equivalently priced AMD hardware.

> They have such a dominant market share because they make more compute for less dollar.

Is that really the case? I looked at getting an accelerator card for personal use. AMD’s Radeon Instinct MI25 will run me $900 - $1000. A similarly specced NVidia Tesla P100 is in the $3000 range.

They don't make more compute for less. They just win via cuda. AMD is an afterthought. Well I guess NVIDIA is not the only one to blame. The ML world basically sold their soul to the level they decided to use cuda.

I'm not sure where the limitation actually was. I've been running RTX Voice on a GTX-970 for a number of months and didn't have to configure anything special other than being told to install "RTX Voice" for non RTX cards and "Nvidia Broadcast" on the newer RTX cards.

The noise cancelling itself is dark magic though - and I especially like the fact it'll work with both input and output - eg it'll clean up your microphone but also do noise reduction on incoming streams too. For me, it's been great in voice chat when other people have rubbish microphones, and has made a number of inaudible youtube videos easy and clear to understand.

>I'm not sure where the limitation actually was.

Simple, the sales and marketing department.

> Simple, the sales and marketing department.

Sure, though I think the point I was trying to make is more "I've heard about this limitation, but it's not something I've actually encountered even with lots of usage".

That's not quite true. Having tensor cores available makes ML algorithms like this more efficient which means lower GPU utilization and lower power draw.

Ultimately even a 9xx is going to have plenty of headroom for an algorithm like this, so it doesn't make sense to lock it off - but there is a difference. I understand the decision to lock it off for a combination of market segmentation (convince people to upgrade) + ensuring the people who use it get the best experience (0 degradation in framerate, near-0 increase in power draw)

I assumed the limit had more to do with testing surface area than product segmentation. I can't imagine a significant number of people made the jump to RTX based solely on the voice features.

If you are on Linux it’s worth checking out PulseEffects [0]. Setting up gate, limiter, etc. is a great way to tune your audio and in a more robust way than some AI algorithm that only runs on some hardware.

[0]: https://github.com/wwmm/pulseeffects

For someone not at all familiar with tuning audio profiles, do you know of any recommended starting point for tuning out background noise for calls using PulseEffects?

Sure check out these two articles, one about tuning it for better sound output [0] and the other for tuning the microphone [1]. Hope that can be a good starting point.

[0]: https://fedoramagazine.org/tune-up-your-sound-with-pulseeffe... [1]: https://fedoramagazine.org/tune-up-your-sound-with-pulseeffe...

Seems like it should be possible to have an "auto" button?

Is there any planned support for Linux, I'm aware of Cadmus (and can't get it to produce any noise) & Noise-torch (only turns off the mic when there's non-speech) but neither of these haven't worked to effectively filter out background noises.

Have a look at https://sr.ht/~arsen/alsa_rnnoise/. It does noise cancellation as an Alsa plugin and in my tests effectively filtered out background noises. I have it running permanently since then and observed no drawback.

Nice! I've been using NoiseTorch, which works great but has as major downside I can't configure it to autoload, so I have to remember to set it myself after boot. There's also PulseEffects, which supposedly has an RNN plugin, but it didn't work for me when I tried it some months ago.

Does Alsa_rnnoise work well with Pulse/Pipewire systems?

(NoiseTorch maintainer here)

This is a limitation that will likely be lifted with the pipewire version. Pipewire 0.3.25 released yesterday which added some of the things we require, still some other things missing, but it seems like things are happening.

In the meantime it is possible to configure NoiseTorch as a systemd service[0], but it's not recommended because it constantly uses CPU even if no application is actually using the microphone. Afaik PipeWire uses a pull model where it pulls data when it needs it, so it shouldn't use CPU to denoise the microphone when it's not in use.

So it is currently possible to do that, but I'm deliberately not making it easy via the UI, because I don't want to deal with complaints about NoiseTorch killing people's laptop batteries or whatever.

[0]: https://github.com/lawl/NoiseTorch/wiki/Start-automatically-...

Thanks for your reply! As I use NoiseTorch, it already knows the previous configuration, which is always what I want, so the only thing i need to do is press 'load'. I risk being terribly naive, but it seems like a flag (`noisetorch --autoload` or `--loadlastconfig`) could solve this as well. Then I could just run it that way on session start. Would that be possible?

Should be possible. If someone wants to implement that, i suggest opening an issue with an implementation proposal and then send a PR.

> Does Alsa_rnnoise work well with Pulse/Pipewire systems?

I don't think so. I think that Pipewire and Pulseaudio would like to replace that part of Alsa, and I see no documentation on how to mix Alsa plugins like this (that work with the sound) with their Alsa plugins (that transports Alsa sound to the soundservers).

I would recommend PulseEffects. I switched from NoiseTorch and haven't looked back. It is a much more polished UI and has other effects if you need them. It also automatically stops doing audio processing when no apps are using the output and can easily be configured to auto-start so it can be basically set it and forget it.

I wish there was something similar on Macs available for free. I have seen a few that do this but they all had subscription fees attached. I just need something to muffle my keyboard when I'm taking notes on calls, and muting/unmuting gets very tedious.

I use this https://krisp.ai/. There is a free tier, but it’s cheap at 60/year given that works pays for it. It’s been great.

It works on PC and Macs. You can filter out noise on your end or the others persons end. I highly recommend it.

I have no horse in this race, but I’d like to push back on a $60 annual subscription for what’s essentially a static piece of software as ‘cheap.’ Since when did everything have to be paid for on a subscription basis?

This is nothing new, "static" software with yearly licenses has always been around and if anything it's a guarantee that bugs will get fixed, features will get added and your data won't be sold to third parties.

For a business this is basically free. You could honestly even think of it as a net-positive for some teams. Think of all of the times you were in a group meeting and somebody has to repeat themselves because someone else was typing, living next door to construction, ambulance drove by, etc. $300 / year for a team of 5 more than pays up for itself if you think of it in terms of hourly salary. For a $120k/year salaried engineer, effectively paid $60 / hour, you only need to save them 1 hour per year for this purchase to be financially worth it for the business. Not to mention the less tangible benefits like "being less distracted/stressed during meetings".

Well, it's not particularly static. I get regular updates to the ML model (which is where the vast majority of the value comes in), and other updates that have been useful (the ability to sync mute status between hardware and software, so you don't get the infamous "double mute").

I tried krisp a year ago but aside from the price, if I'm remembering correctly the thing that scared me off of it was that it seemed to render japanese speech unintelligible, as if it had only been trained on english. The nvidia one is the best but the open source ones are probably also better than krisp.

I looked at Krisp but I felt the cost was a little high as work is not going to cover that expense for me.

muting/unmuting gets very tedious.

This is a failure of meeting software. The #1 critical feature of the $$$$$ conferencing systems that get installed into e.g. boardrooms is a mute button that activates instantly, shows its state, is easy to reach, and never fails.

A push-to-talk shortcut can be a lot easier than switching apps back and forth and trying to click on the tiny microphone icon.

I have a very noisy environment, but after reading up on the topic I realized I don't need noise cancellation anymore. All I needed was a properly setup software-based equalizer that filters out noise regions like electricity and fans.

Noise cancellation, especially ML based ones, have weird artifacts which I assume result from false positives. Sure, you can filter out a hair dryer or a vacuum cleaner with those and still get a "decent" result. But if you have things like that constantly in your background probably you shouldn't try to record any voice there in the first place ...

It's not a always a choice. The rain on my roof or a bus outside will cause significant noise which can't be EQ'd out. If you don't need it, that's fine. But not everyone can find perfect conditions for recording / calls.

One thing I really like about the ML option is how simple it is to get other people to use it. I got a friend to install it and use it in just a few of minutes because I was tired of listening to his MX Blue keyboard.

For actual audio recordings I would never use something like this, but for voice chat it's perfect because it allows a person to have a good enough level of audio quality with minimal effort.

This, plus using a higher quality microphone than the ones that are integrated in webcams and notebooks. Since last year, I'm using a Snowball iCE that is mounted fairly close to my mouth. This not only improved sound quality but also reduced background noise quite a bit.

Make sure to use a gate and/or compressor along with the EQ. If you just use an EQ you will sometimes get transient glitches that it can't easily or effectively filter out by itself (I had lots of problems with this)

Ah, nice. Just in time for summer. Less AC noise on stream is good for everybody.

Reply on news.ycombinator.com