8 ms·

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

109 points by woliveirajr 6y ago

Jiejeing 6y ago
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.
- sto_hristo 6y ago
  If there was strong competition, none of that circus would be possible.
  - willis936 6y ago
    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.
    - kevingadd 6y ago
      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.
    - 10000truths 6y ago
      > 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.
    - rowanG077 6y ago
      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.
- BuildTheRobots 6y ago
  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.
  - delfinom 6y ago
    >I'm not sure where the limitation actually was. Simple, the sales and marketing department.
    - BuildTheRobots 6y ago
      > 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".
- kevingadd 6y ago
  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)
- tengbretson 6y ago
  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.
Causality1 6y ago
Ah, nice. Just in time for summer. Less AC noise on stream is good for everybody.
everyone 6y ago
Sure would be nice if I could actually get a GPU to actually play games with, and not just fucking waste electricity for no reason.
nannal 6y ago
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.
- onli 6y ago
  Have a look at https://sr.ht/~arsen/alsa_rnnoise/ 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.
  - brnt 6y ago
    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?
    - lawl 6y ago
      (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-with-Systemd https://github.com/lawl/NoiseTorch/wiki/Start-automatically-...
      - 6y ago
numlock86 6y ago
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 ...
- viraptor 6y ago
  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.
- innodendo 6y ago
  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.
- EveYoung 6y ago
  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.
- kevingadd 6y ago
  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)
ruph123 6y ago
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 https://github.com/wwmm/pulseeffects
- Haegin 6y ago
  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?
  - ruph123 6y ago
    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-pulseeffects-speakers/ https://fedoramagazine.org/tune-up-your-sound-with-pulseeffe... [1]: https://fedoramagazine.org/tune-up-your-sound-with-pulseeffects-microphones/ https://fedoramagazine.org/tune-up-your-sound-with-pulseeffe...
    - neolog 6y ago
      Seems like it should be possible to have an "auto" button?
sylens 6y ago
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.
- janesvilleseo 6y ago
  I use this https://krisp.ai/ 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.
  - patentatt 6y ago
    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?
    - steinuil 6y ago
      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.
    - Sodman 6y ago
      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".
    - 6y ago
NikolaNovak 6y ago
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). :-/
- patentatt 6y ago
  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.
  - NikolaNovak 6y ago
    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 :-(
  - kQq9oHeAz6wLLS 6y ago
    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.

109 points by woliveirajr 6y ago

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.

sto_hristo 6y ago
If there was strong competition, none of that circus would be possible.
- willis936 6y ago
  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.
  - kevingadd 6y ago
    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.
  - 10000truths 6y ago
    > 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.
  - rowanG077 6y ago
    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.
BuildTheRobots 6y ago
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.
- delfinom 6y ago
  >I'm not sure where the limitation actually was. Simple, the sales and marketing department.
  - BuildTheRobots 6y ago
    > 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".
kevingadd 6y ago
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)
tengbretson 6y ago
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.

Causality1 6y ago

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

everyone 6y ago

Sure would be nice if I could actually get a GPU to actually play games with, and not just fucking waste electricity for no reason.

nannal 6y ago

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.

onli 6y ago
Have a look at https://sr.ht/~arsen/alsa_rnnoise/ 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.
- brnt 6y ago
  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?
  - lawl 6y ago
    (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-with-Systemd https://github.com/lawl/NoiseTorch/wiki/Start-automatically-...
    - 6y ago

numlock86 6y ago

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 ...

viraptor 6y ago
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.
innodendo 6y ago
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.
EveYoung 6y ago
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.
kevingadd 6y ago
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)

ruph123 6y ago

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 https://github.com/wwmm/pulseeffects

Haegin 6y ago
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?
- ruph123 6y ago
  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-pulseeffects-speakers/ https://fedoramagazine.org/tune-up-your-sound-with-pulseeffe... [1]: https://fedoramagazine.org/tune-up-your-sound-with-pulseeffects-microphones/ https://fedoramagazine.org/tune-up-your-sound-with-pulseeffe...
  - neolog 6y ago
    Seems like it should be possible to have an "auto" button?

sylens 6y ago

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.

janesvilleseo 6y ago
I use this https://krisp.ai/ 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.
- patentatt 6y ago
  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?
  - steinuil 6y ago
    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.
  - Sodman 6y ago
    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".
  - 6y ago

NikolaNovak 6y ago

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). :-/

patentatt 6y ago
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.
- NikolaNovak 6y ago
  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 :-(
- kQq9oHeAz6wLLS 6y ago
  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.