Noise Cancellation for Remote Workers: The Three-Layer Stack

It’s 3:18 PM on a Wednesday and a vendor truck is rolling past my window announcing tomatoes through a loudspeaker bolted to its roof. The maid is on the phone in the kitchen having an animated argument with someone about a delivery. A rooster is shouting at nothing. Reggaeton is playing from a balcony three houses over, the bass loud enough that I can feel it in the desk. There’s a tin roof somewhere within fifty metres and the daily 3 PM rain is starting to hit it.

I’m on a video call with the head of finance at my company. Camera live. Mic live.

She is asking me a question about Q3 forecasting and the only thing she can hear is my voice answering it.

This is the noise cancellation stack. It’s three layers running at the same time, each one catching what the layer before it missed, plus a proper headset feeding clean audio into the front of the chain. None of it is exotic. Most of it is free if you have the right hardware. And once it’s set up, it just works, every day, while the world around you gets on with whatever the world around you is doing.

Noise is the thing that gets you caught. Not your IP. Not your time zone. Not a window. The dog that loses its mind when the manager asks you a direct question.

Article 02 / The hardware that runs all this The Best Workstation Laptops for Remote Work Anywhere The three-machine setup, the workstation GPU rule, and why a used ThinkPad P15 with an RTX A4000 is the cheapest path to a free Nvidia Broadcast install. ← The Premise

Why one layer is never enough

The instinct most people have is to install one noise cancellation app and consider the problem solved. That works fine in a Toronto suburb where the worst noise is a neighbour’s lawn mower on Saturday morning. It does not work in a place where the noise is sustained, varied, and often loud. It does not work in wet season. It does not work during a Quinceañera. It does not work when the maid decides to vacuum during a client demo and the parrot next door starts screaming at her through the wall.

Each noise suppression system has a model of what voice looks like and what isn’t voice. That model has limits. Push it past those limits with sustained loud noise and it starts producing artefacts. Robotic-sounding voice processing. Brief dropouts. Moments where the system can’t decide if the sound it’s hearing is you talking or a vendor truck rolling past.

Run two layers in sequence and what one lets through, the next catches. Run three and the combined result is genuinely robust. The first time you take a call during a downpour on a tin roof and nobody comments, you understand why. The chain matters more than any single link in it.

Layer 01 / The Microphone Itself

Hardware processing on the boom

The first layer doesn’t run in any app on your computer. It runs on the headset. Specifically, on the dedicated DSP chip in the boom microphone of a professional headset, processing audio before it reaches your computer at all.

The recommendation here is the Jabra Evolve2 line. These are not consumer headsets with marketing copy about “crystal clear audio” on the box. They’re professional communication headsets built for sustained daily use in busy office environments, which in noise terms maps almost exactly to what a remote worker in a tropical city deals with day to day. The Evolve2 line has hardware DSP noise cancellation built into the boom mic, independent of any software. The mic picks up your voice at close range, the DSP suppresses ambient noise, and a clean signal goes into the computer before any software touches it.

The boom positioning matters more than most people appreciate. The mic element sits two centimetres from your mouth and picks up your voice at close range, which dramatically increases the signal-to-noise ratio before any processing happens. A laptop’s built-in mic array, however good, is picking up your voice from 30 to 50 centimetres away. It’s also picking up the vendor truck, the parrot, and the rain on the tin roof at roughly the same intensity. Starting with a cleaner input makes every software layer downstream of it more effective.

The Evolve2 lineup

• Evolve2 40. The entry wired option. Single or dual ear, USB-A or USB-C, excellent boom mic, comfortable for all-day wear. The right pick if budget is the constraint and you don’t need wireless.
• Evolve2 55. Mid-range wireless. 14-hour battery, active noise cancellation in the earcups for listening comfort, the same boom-mic DSP as the higher tiers. Sweet spot of the range for most remote workers.
• Evolve2 75. Flagship. 36-hour battery, stronger earcup ANC, premium build. For users in noticeably louder environments who are on calls all day, the comfort difference over an eight-hour stretch is real.

All three work with any call platform. None of them require proprietary software for basic function. Plug it in, select it as your audio device, done.

Layer 02 / Software on the Call Machine

The free option most people miss

The second layer runs on the call machine (the laptop dedicated to the all-day Google Meet) and handles software noise removal before your audio reaches the meeting. What you use here depends entirely on what GPU is in that machine. Most articles about remote work audio default to recommending Krisp because Krisp pays for affiliate placement. The reality is that if your machine has a modern dedicated Nvidia or AMD GPU, this layer costs you nothing.

Modern Nvidia RTX: Nvidia Broadcast (free)

Nvidia Broadcast is the right pick for any RTX-series card and it’s entirely free. It uses the GPU’s dedicated tensor cores for real-time AI noise removal, which means the processing load goes on the GPU and not the CPU. Quality is genuinely excellent. Handles fans, rain, street noise, and sustained ambient noise without introducing audible artefacts. Comes with virtual backgrounds and webcam effects baked in, which I’ll come back to in a minute. Install Broadcast, set it as the microphone input in Google Meet, done.

The same logic applies to any used workstation laptop with an RTX A-series card from Article 2. RTX A2000, A3000, A4000, A5000, the new RTX 2000/3000/4000 Ada series. All RTX, all Broadcast-compatible, all free.

Older Nvidia GTX: RTX Voice (free)

RTX Voice is the legacy free option for older Nvidia cards. Officially supported on the GTX 16-series and above, with a community workaround that extends it to some GTX 10-series cards. It does the same core noise suppression as Broadcast without the additional virtual background and webcam features. If your machine has an older Nvidia card that can’t run Broadcast, RTX Voice is the right answer. Quality is very good for any normal remote work use case.

Modern AMD: AMD Noise Suppression (free)

AMD Noise Suppression is built directly into the AMD Software: Adrenalin Edition driver. No third-party install. Uses the GPU’s AI acceleration on RDNA 2 and RDNA 3 cards, integrates at the system level as a virtual microphone selectable in any application. Performance on the modern AMD cards is genuinely competitive with Nvidia Broadcast. Open Adrenalin, enable Noise Suppression, select it as your mic input in Meet. Free, no subscription, immediate.

Intel integrated graphics, no GPU, or Mac: Krisp (paid)

If your call machine has Intel integrated graphics, no dedicated GPU, or you’re running a MacBook, Krisp is the right software choice. Runs entirely on CPU, which means it works on any machine regardless of graphics, and CPU usage is low on modern hardware.

What makes Krisp particularly useful for the call-machine role is that it works in both directions at the same time. It cleans your outgoing mic feed going into the call. It also cleans the incoming audio coming from the other end. So if a colleague joins the call from a noisy environment with nothing running on their side, Krisp removes their background noise on your side. You hear a clean conversation even when their setup is bad. No other tool in this list does that.

The free tier is capped at a number of minutes per week. For full-time all-day use you need the paid subscription. Treat it as professional infrastructure, the same category as your internet.

Layer 03 / Software on the Workstation

Same logic, different machine

The third layer runs on the primary workstation and handles noise suppression for the softphone (the VoIP application used for direct work calls separate from the all-day Meet). The same GPU rules apply.

• Modern Nvidia RTX or RTX A-series: Nvidia Broadcast (free).
• Older Nvidia GTX: RTX Voice (free).
• Modern AMD RDNA 2 / RDNA 3: AMD Noise Suppression in Adrenalin (free).
• Intel integrated, no dedicated GPU, or Mac: Krisp (paid).

If you followed the workstation recommendations in Article 2 and are running a Dell Precision, HP ZBook, or ThinkPad P-series, you almost certainly have an Nvidia RTX A-series GPU. Broadcast is your free answer. The total software cost across both machines for someone running the recommended used workstation setup is zero.

The Open-Source Option

RNNoise for the technically inclined

If you’d rather run entirely free and open-source software and you’re comfortable with some configuration, there’s a real alternative to every paid tool above: noise-suppression-for-voice by werman on GitHub. Built on Xiph’s RNNoise, over 6,000 GitHub stars, actively maintained.

The plugin ships in VST2, VST3, LV2, LADSPA, AU, and AUv3 formats and is designed to suppress fan noise, office noise, crowds, traffic, rain, and construction sounds. It works at the system level, so once it’s set up it applies to any application that uses your microphone, not just specific call apps.

Setup on Windows

Two pieces of software working together.

1. Equalizer APO. A free, open-source system-wide equalizer for Windows that hosts VST plugins. Get it from equalizerapo.com. During install, select only your capture (input/microphone) devices, not playback devices. Reboot afterwards.
2. The RNNoise VST plugin. Grab the latest release from the GitHub releases page, unzip, and copy bin/vst/librnnoise_vst.dll to a stable location on your C drive. Open Equalizer APO’s editor, add a VST Plugin control, point it at the DLL. Noise suppression is now active system-wide on the selected microphone.

The VAD (Voice Activity Detection) threshold is configurable in the plugin, which lets you tune how aggressively it separates voice from background noise. With the threshold dialled in correctly, quality is genuinely comparable to paid software. Fan noise, keyboard clatter, rain, road noise, vendor trucks, the lot.

Setup on Linux

Linux is arguably cleaner than Windows for this. PipeWire users on version 0.3.45 and above can drop the plugin in via a config file. PulseAudio users load it as a module directly. If you want a GUI rather than manual config, EasyEffects and NoiseTorch both wrap the same RNNoise engine and are in most distro package managers.

On macOS

The plugin technically supports AU and AUv3 but the macOS implementation is flagged as a work in progress in the repo. Mac users are better off with Krisp.

The trade-off with the open-source path is straightforward. More setup effort up front, more comfort with audio routing required, no recurring cost ever. For someone technically inclined who doesn’t want subscriptions and is on Windows or Linux, it’s a real option. For everyone else, the GPU-native free tools or Krisp get you to the same outcome with less typing.

The Final Pass

Google Meet’s built-in suppression

Google Meet has noise suppression built in and it’s on by default. In the layered stack it runs as a final pass: after the headset DSP, after your GPU-based or Krisp software layer, catching anything residual. Keep it enabled. It’s free, it adds a final cleanup pass on top of everything else, and it costs nothing in performance.

On its own, Meet’s suppression is not enough for sustained heavy noise. It handles light intermittent background fine. It struggles with the consistent loud noise of a tropical environment in wet season, or any room with running fans, or a busy street outside an open window. This is exactly why a dedicated software layer always runs before it in the chain.

The Visual Layer

Virtual backgrounds, always on

Noise cancellation handles audio. Virtual backgrounds handle the visual side. Both matter equally for staying invisible.

A virtual background running at all times means your physical surroundings are never on the call. That removes an entire category of problems in one move. Wrong time of day visible through a window. Something identifiable behind you. The maid walking through a screen-share. A delivery showing up at the door. Lighting changes across the day that make it obvious you’re not where work thinks you are.

For Google Meet:

• Use a static professional image, not a video background. Video backgrounds add unnecessary processing load on the call machine and edge detection is worse on them.
• Light yourself from in front, not behind. Backlighting wrecks edge detection and you end up with cut-out artefacts around your hair and shoulders. A window in front of you, or a cheap key light, is enough.
• Sit in a consistent position in the frame. Virtual background edge detection works best when you’re where it expects you to be.

On a dedicated call machine running only Google Meet, the virtual background runs with no real performance impact because nothing else is competing for resources. This is one of the practical advantages of giving the call its own machine. Nothing else on it is asking the GPU for anything.

The Trick That Sounds Strange Until You Try It

Headset on one ear, AirPod on the other

Running two machines at the same time, one for actual work and one for the all-day Meet, creates a small problem. You need to hear both at once without switching between them or missing something.

The fix is two audio devices on simultaneously. Jabra on the left ear in single-ear configuration, handling all softphone and direct work-call audio plus the boom mic. A single USB-C AirPod in the right ear, paired to the dedicated call machine, feeding the ongoing Google Meet audio.

The two streams are completely independent. The Jabra rings when a softphone call comes in. The Meet audio sits at low ambient volume in the right ear so you can hear someone say your name without staring at the call machine. After the first afternoon you stop noticing.

One ear open also keeps you aware of the immediate physical environment, which is genuinely useful. You hear when the maid is about to come in. You hear the delivery truck before it knocks. You hear the rain start before it gets loud enough to need to mention.

Before You Trust It

Test the stack under real conditions

The worst way to find out the stack isn’t working is on a live call with your manager. Test it before you trust it.

The straightforward test:

1. Open Google Meet and start a preview. Meet shows you exactly what your audio and video look like before joining a real call.
2. Run the full noise environment you typically work in. Fans on. Windows open if your environment is noisy. Whatever conditions you actually deal with.
3. Have someone join a test call from a different machine and report what they hear. Self-assessment is unreliable. The other end of the line is the only honest judge.

Krisp’s built-in test tool is underrated. It records a short clip with Krisp on and off and lets you A/B them. Run it in your actual working conditions, not a quiet test, and you’ll hear exactly what gets through.

Signs the stack needs adjustment

• Robotic or clipped voice quality. Usually means the noise input is overwhelming a layer somewhere. Check microphone input levels first; a hot input gives the suppression tool less headroom to work with.
• Intermittent voice dropouts. Often a CPU or GPU load issue. Check what else is competing for resources during the call. If the call machine has a Chrome window with twenty tabs open, close them.
• Noise audible during pauses but not during speech. Voice activity detection threshold is wrong. Check the sensitivity setting in Krisp or Broadcast. The fix is usually moving it more aggressive, not less.

All Together

The full stack at a glance

Layer	Nvidia RTX	Nvidia GTX	AMD RDNA 2/3	Intel iGPU	Mac
Headset DSP	Jabra Evolve2	Jabra Evolve2	Jabra Evolve2	Jabra Evolve2	Jabra Evolve2
Call machine	Nvidia Broadcast (free)	RTX Voice (free)	AMD Noise Suppression (free)	Krisp (paid)	Krisp (paid)
Workstation	Nvidia Broadcast (free)	RTX Voice (free)	AMD Noise Suppression (free)	Krisp (paid)	Krisp (paid)
Final pass	Google Meet built-in	Google Meet built-in	Google Meet built-in	Google Meet built-in	Google Meet built-in
Open-source alt	RNNoise + EQ APO	RNNoise + EQ APO	RNNoise + EQ APO	RNNoise + EQ APO	Limited support

The whole audio stack at a glance

The room itself

One last thing, and the cheapest part of the whole article. Software and hardware do most of the work, but the room you’re sitting in does the rest. There are three things that close the gap between “this works most of the time” and “this never fails.”

Close the door. A closed door takes about 15 dB off whatever’s happening on the other side of it. Closing one solid wood door is the single biggest improvement-per-effort move available to a remote worker. Costs nothing. Takes one second.

Soft surfaces in the room. Hard tile and bare walls reflect sound and make every echo louder. A rug, curtains, a couch, a bookshelf full of books, anything that absorbs sound rather than bouncing it, all reduce ambient noise pickup measurably. You don’t need acoustic foam panels. You need a normally-furnished room.

Face the window. Two things at once. The window becomes the light source on your face, which makes virtual background edge detection cleaner. And you’re facing away from the room, which means the directional element of the boom mic is pointing at less of it.

Three free moves. Combined effect bigger than people expect.

Software and silicon do most of the work. A closed door does about 15 dB more.

Article 04 / Up next Power Redundancy for Remote Workers UPS sizing, the DC UPS for your router that almost no one talks about, USB-C battery banks, and how to run an entire workday with no mains power. → Up Next / Article 04 Power Redundancy for Remote Workers →