- Top Pick: Samson Q2U
- Runner-up: Shure MV7+
- Also consider: Elgato Wave:3
A mechanical keyboard, a case fan and an open window are all it takes to make your callouts the least welcome sound in your squad’s voice channel. Chat apps hear everything the microphone hears, and most gaming microphones catch all of it. You can treat the room, buy a quieter keyboard, or fix the signal in software — and on Linux, the small open-source tool built for exactly that job is NoiseTorch.
NoiseTorch, maintained today as NoiseTorch-ng, loads a noise suppression module into your PulseAudio or PipeWire server and creates a virtual microphone that any game or chat app can select like a hardware device. This guide explains how that virtual microphone works, what its voice activation threshold does, how to install it the documented way, where its honest limits are, and which microphones pair best with a filtered setup.
As an Amazon Associate, we earn from qualifying purchases.
Quick answer
NoiseTorch-ng is a free, open-source Linux app that creates a virtual microphone cleaned up by the RNNoise library. Pick your real microphone, click Load, then choose Filtered Microphone inside Discord, Mumble or your game. A threshold slider decides how strictly non-voice sound is muted, the app can be closed once loaded, and documented added latency on the microphone path is 10 milliseconds.
What NoiseTorch actually does
NoiseTorch does not patch individual apps. When you press Load, it loads modules into your running audio server — PulseAudio or PipeWire — and those modules register a new input device named Filtered Microphone. Your real microphone feeds into a LADSPA filter built on RNNoise, the recurrent neural network suppressor developed at Xiph.org and Mozilla, and whatever survives the filter is what the virtual device passes on. Because the result is an ordinary system audio device, it works in Discord, Mumble, OBS and in-game voice chat alike, with no per-app configuration.
The same trick works in reverse for output: route application audio to the Filtered Headphones device and a teammate’s noisy microphone gets suppressed for you. One documented detail matters for daily use: the window does not need to stay open — once loaded, the virtual microphone keeps working until you explicitly unload it, and the app process itself is no longer required.
The voice activation threshold, explained
RNNoise judges every small frame of audio and estimates how likely it is to contain voice. NoiseTorch turns that estimate into a gate with one control: the Voice Activation Threshold slider, which runs from 0 up to a maximum of 95. Frames that score below your threshold are muted; frames above it pass through, still denoised. The documentation advises running the threshold as high as your microphone allows — with a decent microphone, the maximum is realistic — and lowering it gradually if your speech starts cutting out.
At a threshold of 0 the gate never closes, but the dampening still applies, which suits players who would rather keep a faint fan than lose the first syllable of a quiet callout. One quirk to know: threshold changes only apply after you unload and reload the virtual microphone, so tuning is a short cycle of adjust, reload, and a mic check in your chat app.
Installing NoiseTorch the documented way
The documented route is the release archive from the project’s GitHub releases page. Unpack the tarball into your home directory and it places the binary, icon and desktop entry in the right locations. One extra step is required: grant the binary the CAP_SYS_RESOURCE capability with setcap, which lets it request the scheduling priority it needs, and make sure your local bin folder is on your PATH. Building from source is the other documented route, using a Go compiler and a C compiler with make.
After a past security scare around the original update server, the community audited the code, reported no malicious code was found, and updates now come from the releases page — another reason to install from there rather than a random mirror. Day-to-day use is then deliberately boring: open the app, select your microphone, click Load, and choose Filtered Microphone in your game or chat app. Click Unload when you want your raw devices back.
NoiseTorch vs EasyEffects vs RTX Voice
The closest relative is EasyEffects, which the NoiseTorch documentation itself names as an alternative. EasyEffects is a full PipeWire effects suite — equalizers, compressors, echo cancellation and an RNNoise plugin among many others — and it only supports PipeWire. It offers far more control, but it is a chain you configure and keep running, where NoiseTorch is one job done in two clicks, works on PulseAudio as well, and can be closed after loading.
On Windows, NVIDIA’s RTX Voice and Broadcast noise removal do a similar job on the GPU, but they are proprietary Windows software tied to NVIDIA hardware, with no Linux version. For a Linux gamer, RNNoise through NoiseTorch or EasyEffects is the practical equivalent: a neural network separating voice from steady noise, running on the CPU inside your audio server instead of on a graphics card.
The honest limits
NoiseTorch filters one microphone at a time: you select a single input device to denoise, so a two-microphone streaming setup is outside its design. It is built for steady background noise — fans, keyboard clatter, traffic hum, a passing train. It cannot rescue a genuinely loud room, because a television or a conversation behind you contains actual voices, and a voice detector has no reason to mute them. A microphone across the room defeats it too; suppression works best when your voice is clearly the loudest thing it hears, a placement problem no filter fully solves.
It also does noise suppression only, not echo cancellation, so play through headphones rather than speakers if you do not want game audio looping back. The documentation puts output filtering latency at roughly 100 milliseconds on PulseAudio, fine for chat but worth knowing before you filter your own monitoring path; the input path’s 10 milliseconds is what matters for callouts. Finally, expect an occasional reload if the audio server restarts or your device list changes.
Microphones that pair with a filtered setup
Suppression is the second line of defence. The first is a microphone that hands it a clean, close signal to begin with.
Samson Q2U
A dynamic handheld microphone is the easiest win in a noisy room, because a cardioid dynamic capsule naturally rejects sound that is not close and on-axis. The Q2U connects over USB-C for plug-and-play use or XLR when you later add an interface, and ships with a tripod stand, clip, windscreen and both cables.
Specs:
– Type: Dynamic, cardioid
– Connections: USB-C and XLR
– Digital conversion: 16-bit, 48 kHz
– Frequency response: 50 Hz to 15 kHz
– Monitoring: 3.5 mm headphone output
Shure MV7+
The MV7+ is the premium step up in the same dynamic, dual-output mould: USB-C for today, XLR for the interface you buy later. Its onboard DSP brings an Auto Level Mode and a digital pop filter, and the LED touch panel handles muting on the microphone itself. Shure’s MOTIV Mix desktop app is published for Windows and Mac, so on Linux treat the MV7+ as a class-compliant USB or XLR microphone.
Specs:
– Type: Dynamic, cardioid
– Connections: USB-C and XLR
– DSP: Auto Level Mode, digital pop filter, reverb
– Controls: LED touch panel with mute
– Model: MV7+-K-BNDL bundle with desktop stand
Elgato Wave:3
A condenser hears more of the room than a dynamic, which makes the Wave:3 the microphone in this list that gains the most from a NoiseTorch filter in front of it. Its strengths are detail and headroom, with Clipguard technology protecting loud moments from distortion. Elgato’s Wave Link mixer targets Windows and Mac; on Linux the Wave:3 simply appears as a USB audio device, ready to be filtered.
Specs:
– Capsule: 17 mm electret condenser, cardioid
– Resolution: 24-bit
– Sample rates: 48 and 96 kHz
– Frequency response: 70 Hz to 20 kHz
– Max SPL: 120 dB, 140 dB with Clipguard engaged
HyperX QuadCast S
The QuadCast S is the gaming-desk option: a USB condenser with tap-to-mute, a gain knob and four selectable polar patterns. Use cardioid for solo play so the microphone favours your voice over the room, then let NoiseTorch handle the keyboard it still hears. HyperX’s NGENUITY software is a Windows app, but the controls that matter here are physical.
Specs:
– Type: USB condenser, tri-capsule array
– Polar patterns: Cardioid, omnidirectional, bidirectional, stereo
– Controls: Tap-to-mute, gain control dial
– Mount: Anti-vibration shock mount, built-in pop filter
– Monitoring: Headphone output
Focusrite Scarlett 2i2 (4th Gen)
If your microphone is XLR-only, or your USB microphone’s XLR output is calling, the Scarlett 2i2 is the interface that feeds it into the chain NoiseTorch filters. The fourth-generation model offers two microphone inputs with an Air presence mode, Auto Gain and Clip Safe features. Focusrite publishes its drivers and Control software for Windows and Mac; on Linux it works as a standard class-compliant USB audio device, and its input is what you select inside NoiseTorch.
Specs:
– Inputs: 2 XLR and line inputs
– Outputs: 2 balanced line outputs, headphone output
– Features: Air mode, Auto Gain, Clip Safe
– Connection: USB-C, bus powered
How to choose
Start with the microphone, not the filter: a close-talked dynamic in a loud room beats an expensive condenser across the desk, with or without suppression. Choose USB for one cable and no extra boxes, and prefer a model with an XLR output as well if an interface is in your future. Then check your audio server — NoiseTorch covers PulseAudio and PipeWire, while EasyEffects is the better detour only if you want a full effects chain. Set the threshold with the mic check in your chat app: raise it until the keyboard disappears, then back it off if your voice clips. Finally, use headphones, because no setting here cancels echo.
Who should skip this
Skip NoiseTorch if your only voice app already has a built-in suppressor you are happy with, since a system-wide virtual microphone is extra machinery for a solved problem. Windows players with NVIDIA graphics already have a GPU-based route. If your room problem is echo, a television, or other people talking, spend the effort on placement and a dynamic capsule first — suppression will not fix those. And if you want mixing, echo cancellation and per-app effects on PipeWire, EasyEffects is the tool you actually want.
FAQ
Does NoiseTorch work on PipeWire, or only PulseAudio?
Both. The project supports PulseAudio and PipeWire, and either way you get the same Filtered Microphone device to select in your apps.
Do I have to keep the NoiseTorch window open while I play?
No. Once the modules are loaded, the documentation says the virtual microphone keeps working until you unload it, and the app process itself is not required anymore. Close the window and play.
Why does my voice cut out at the start of sentences?
Your threshold is too strict for your microphone and speaking volume. Lower the Voice Activation Threshold gradually until quiet speech survives, remembering that every change needs an unload and reload before it applies.
Does NoiseTorch add delay to my voice?
The project documents 10 milliseconds of inherent latency for microphone filtering, a non-issue for voice chat. Filtering your output is heavier — on the order of 100 milliseconds with PulseAudio — so think twice before routing your own monitoring through Filtered Headphones.
Final verdict
NoiseTorch does one thing, documents it honestly, and gets out of the way: a two-click RNNoise filter that lives in your audio server and cleans up voice chat in every app at once. It will not redeem a bad room or a badly placed microphone, and EasyEffects remains the deeper PipeWire toolbox, but for a loud keyboard and a fan in a Discord call, loading Filtered Microphone is the simplest fix a Linux player can make today.











