Every major call app now has a noise filter, and most say it uses AI. The filters are good, but each makes choices about what counts as noise, and those choices explain most of the surprises.
Steady noise and sudden noise
Noise comes in two broad kinds. Steady (stationary) noise stays the same over time: a fan, air conditioning, hiss, hum. Changing (non-stationary) noise starts and stops: a dog, a door, a keyboard, traffic.
Classic filters handle the first kind. Audacity’s Noise Reduction learns a profile from a stretch of noise on its own; its manual says it suits constant sounds such as hiss, hum, whine and fan noise, and is not suitable for irregular noise such as traffic or an audience. The gentle settings in call apps aim at steady noise too: Zoom’s Low level blocks low levels of persistent background noise, and Teams on iPhone says the same of its Low level, naming a computer fan or air conditioner.
Machine-learning filters go after the second kind as well. Here is what each vendor names:
| App | Noises the vendor names | Other people’s voices |
|---|---|---|
| Discord (Krisp) | Barking dogs, vacuums, slamming doors | Left in |
| Google Meet | Typing, a closing door, room echo, building work | Left in, including a TV |
| Microsoft Teams | Shuffling papers, slamming doors, barking dogs | Removed with Voice isolation |
| Zoom | Fans and pen tapping, at Medium | Removed at High |
| Webex | Noisy keyboards, doorbells | Removed with Optimize for my voice |
| Xbox party chat | Gamepad clicks, breathing | Not stated |
Other voices are the hard part
The standard modes leave other people’s voices in. Discord says Krisp removes noise that isn’t a human voice, and Google says Meet won’t cancel voices from a TV or people talking.
Removing a second voice takes a separate mode, and each vendor does it differently. Teams’ Voice isolation keeps your voice using a voice profile, which takes about 30 seconds to create and is needed unless you use a headset. Zoom’s High level removes background speech from others, and its Personalized audio isolation builds a voiceprint for crowded rooms. Webex’s Optimize for my voice removes background voices, for open offices and coffee shops. Apple’s Voice Isolation, in FaceTime, phone calls and supported apps, prioritizes your voice and blocks the sounds around you.
Music
Filters can treat music as noise, which is right for a radio in the next room and wrong when you are playing guitar. Most apps let you step back. Google says to turn noise cancellation off if you need other sounds, like music; Xbox says to turn its filter off to keep music in a party; Teams suggests its Low level when you want to play music for others. Webex has a Music mode, and Zoom has Original sound for musicians, which turns off its standard audio filters.
What the filter costs
- Some of your voice. Discord says Krisp can lower voice quality in a quiet room with a good mic, and Microsoft suggests turning Teams’ noise suppression off for high-fidelity mics in a low-noise room, such as a quiet studio.
- Artifacts. When a filter misjudges the noise, it can leave odd sounds behind. Audacity describes them as random bursts of very short tones, sometimes called musical noise.
- Only your side. A filter cleans what your mic sends. Discord points out that Krisp filters only your side, so if your friends haven’t turned theirs on, you still hear their rooms.
- Doubling up. Two filters on one mic do the same job twice. NVIDIA tells Broadcast users to turn off a device’s own noise removal, so the same effect isn’t applied twice.
CPU, GPU, NPU or cloud
The filter has to run somewhere, and where it runs decides what you need:
- CPU, the computer’s main processor. Krisp’s model runs on your device inside Discord, and Discord turns Krisp off when CPU use gets very high. Windows 11’s Voice Clarity uses low-complexity AI models on ordinary x64 and Arm64 processors, with no extra hardware. Zoom notes that its High level may use more CPU.
- GPU, the graphics card. NVIDIA Broadcast runs its AI effects, Noise Removal among them, on the Tensor Cores of GeForce RTX cards, and needs an RTX 2060 or better.
- NPU, a neural processing unit. Windows Studio Effects, including Voice focus, needs one. Voice focus works on smaller NPUs; the full set of effects needs a Copilot+ PC whose NPU handles more than 40 trillion operations per second.
- Cloud. Google says Meet’s noise cancellation can run in the cloud for some Workspace editions and on the phone itself on many Android models, switching between the two.
Measure before you filter
The noise test can record your mic with your browser’s suppression off and then on, and play both back. If the floor is already low with it off, you may not need a filter at all; if it is high, fix the room first and let the filter clean up what is left.