Noise Suppression Software: Remove Background Noise from Your Mic

How to remove background noise from your microphone is one of the most searched audio questions among streamers, remote workers and gamers — and for good reason. Keyboard clatter, case fans, AC units and room echo bleed into your mic more than you think until you hear your own stream recording. This guide walks through exactly how noise suppression software works, which approaches are worth your time, where the technology sits in your audio chain, what it costs in latency, and how to configure it whether you go the free OBS filter route or a dedicated real-time app.

TL;DR

Noise suppression software analyses your mic signal in real time and attenuates everything that does not look like voice.
RNNoise (used in OBS and many free tools) is a solid baseline — roughly 10 ms latency, handles steady-state noise well.
AI/ML-based denoisers go further — they handle keyboard clatter and intermittent noise — at a modest latency cost of 10–30 ms.
Noise suppression must run before voice effects in the chain, not after, or you introduce artefacts the suppressor then fights.
You can set it up for free inside OBS today, or use dedicated software for system-wide coverage that works in Discord, calls and every other app simultaneously.
VoxBooster places noise suppression first in its processing chain automatically, so it works correctly whether you are changing your voice or just cleaning up a raw mic.

Streaming platforms compress audio heavily. That compression amplifies noise artefacts that sound mild in a raw recording but turn into a constant hiss on Twitch VODs. Discord’s automatic gain control does the same thing — it boosts quiet signals, which means your fan hum gets boosted right alongside your voice. Getting noise suppression right solves both problems at once.

What Noise Suppression Software Actually Does

Noise suppression does not simply apply an EQ cut or a noise gate. A gate silences audio below a threshold — useful for muting completely between sentences, but useless for noise that overlaps your voice in frequency. EQ cuts fixed frequency ranges, which would also cut voice harmonics in the same bands.

Modern noise suppression software works differently. It treats audio as a stream of short frames (typically 10–20 ms each), analyses each frame to separate voice content from noise content, and attenuates the noise while passing the voice. The analysis step is where the approaches diverge.

Traditional spectral subtraction

Older tools use spectral subtraction: measure the noise floor during silence, build a noise profile, subtract that profile from every subsequent frame. This works reasonably well for steady-state noise like fan hum or air conditioning that stays at a constant frequency and level. It struggles with noise that changes — keyboard clicks, mouse clicks, someone talking in the background — because the profile measured during silence does not match what arrives mid-sentence.

RNNoise — the practical free baseline

RNNoise is a recurrent neural network-based noise suppressor developed by Mozilla and Jean-Marc Valin. It uses a GRU (gated recurrent unit) network trained on thousands of hours of noisy speech to classify each frequency band as voice or noise on every frame. Because it learned from real recordings, it handles a much broader noise class than spectral subtraction.

RNNoise is open source, has a very small footprint (runs on a single CPU core), adds roughly 10 ms of latency, and is integrated directly into OBS Studio, Chromium (the engine behind Discord’s browser tab) and many other tools. For most people doing casual streaming or calls, RNNoise is genuinely sufficient.

Where RNNoise shows limits: very loud intermittent noise (mechanical keyboard at high WPM right next to the mic), room echo/reverb, and complex acoustic environments with multiple competing sound sources. It also has a fixed model — it cannot be tuned to your specific room without recompiling.

AI/ML denoisers

Newer noise suppression software uses larger neural networks trained on more data, with more parameters dedicated to separating voice from complex backgrounds. Tools like Krisp, NVIDIA RTX Voice (part of NVIDIA Broadcast), and VoxBooster’s noise suppression fall in this category. The typical improvement over RNNoise is most visible in:

Keyboard noise — a 60 WPM typist on a clicky mechanical keyboard right next to the mic, which RNNoise handles poorly.
Room echo and reverb — reflections that arrive milliseconds after the original sound and blur speech intelligibility.
Background speech — someone else talking in the same room, which has voice-like spectral characteristics that confuse simpler models.
Variable-level noise — traffic that surges with each passing vehicle, HVAC that cycles on and off.

The trade-off is compute cost. Heavier models require more CPU (or a GPU for the largest variants). Latency stays in the 10–30 ms range for real-time implementations because they use small chunk sizes, but model inference time sets a floor.

Where Noise Suppression Sits in the Audio Chain

This is where most setups go wrong. The position of the noise suppressor relative to other processing stages matters enormously.

Correct order: suppress first

Microphone → Noise Suppression → Voice Effects / EQ / Compression → Output

If noise suppression runs first, it receives the cleanest possible input: raw mic audio before any processing has added artefacts. It can accurately identify noise versus voice because neither has been altered.

What goes wrong when you reverse it

If you put a pitch-shifter or voice changer before the noise suppressor, the chain looks like:

Microphone → Voice Changer → Noise Suppression → Output

The voice changer creates new harmonics, shifts formants, and produces spectral content that was not in the original recording. The noise suppressor then encounters this modified signal and cannot reliably distinguish voice harmonics (now at unfamiliar frequencies) from noise. The result is voice that gets partially suppressed alongside the noise — thin, watery, or with consonants clipped.

VoxBooster enforces the correct order internally. Its noise suppression stage runs at the top of the processing chain regardless of which other features you have active. This is not an obvious detail until you have heard what the wrong order sounds like.

Gain staging around the suppressor

Noise suppression algorithms are typically calibrated for a specific input level range. If your mic is gain-staged too hot (signal peaking near 0 dBFS), clipping artefacts enter the signal before suppression and cannot be removed. If it is too quiet, the algorithm operates at a low SNR and may not distinguish voice from noise floor reliably. Target a mic level where speech peaks around -12 to -6 dBFS before the suppressor.

RNNoise vs AI/ML: Head-to-Head Comparison

Feature	RNNoise	AI/ML Denoiser (e.g. VoxBooster)
Latency	~10 ms	10–30 ms
CPU usage	Very low (< 5% on 1 core)	Low–moderate (5–20% depending on model)
Steady-state noise (fan, AC)	Excellent	Excellent
Keyboard / click noise	Moderate	Excellent
Room echo / reverb	Poor–moderate	Good–excellent
Background speech	Poor	Good
Variable-level noise	Good	Excellent
Tuneable strength	No (on/off)	Yes (suppression level slider)
Works system-wide	No (per-app)	Yes
Cost	Free	Paid (trial available)

The table makes the decision fairly straightforward. If your acoustic environment is clean with only a steady fan or AC hum, RNNoise in OBS or Discord is all you need — free, no installation. If you have keyboard noise, echo, or need noise suppression to work across every app at once (not just OBS), a dedicated tool earns its cost.

Setting Up Noise Suppression in OBS

OBS Studio has built-in noise suppression via its audio filters. Here is the step-by-step for both RNNoise and the Speex denoiser (a DSP-based alternative):

Add a noise suppression filter in OBS

Open OBS Studio. In the Audio Mixer panel, find your microphone source.
Click the gear icon next to the mic source and select Filters.
In the Filters window, click + (Add) → Noise Suppression.
Give it a name (e.g. “RNNoise”) and click OK.
In the filter settings, choose RNNoise from the Method dropdown.
Click Close.

That is the entire setup. OBS applies the filter in real time to the mic before sending audio to your stream or recording. There is no strength slider for RNNoise — it is on or off. The Speex method has a Suppression Level slider (dB) if you want a tunable option.

Correct filter order in OBS

OBS processes filters top-to-bottom in the Filters window. Noise suppression should be at the top of the list — above any compression, EQ, or voice effects plugins. Drag it to position 1 if needed.

According to the OBS documentation on audio filters, filters run sequentially in the order listed, which is why position matters.

Limitations of the OBS-only approach

The OBS filter only applies to audio processed by OBS. Your Discord mic, browser tabs, game voice chat, and any other application not routed through OBS will not benefit. If you want system-wide noise suppression, you need software that inserts itself at the Windows audio driver level — which is what tools like VoxBooster, Krisp, and NVIDIA Broadcast do.

Setting Up System-Wide Noise Suppression with Dedicated Software

System-wide noise suppression software registers a virtual microphone in Windows. You switch your apps (Discord, OBS, Zoom, your game) to use this virtual mic, and the software handles capturing from your real mic, suppressing noise, and feeding the clean signal to whatever app is listening.

General workflow

Install the software and launch it.
In the software’s settings, confirm your physical microphone is selected as the input.
Enable noise suppression (and set the strength level if the option exists).
In Discord, OBS, Zoom or your game, go to the audio/mic settings and select the virtual microphone the software created as your input device.
Do a test recording to confirm the clean signal is coming through.

VoxBooster setup

Download VoxBooster and install it. The onboarding wizard asks you to select your physical mic and sets up the virtual device automatically. Noise suppression is enabled by default and runs before any voice effects in the chain.

In VoxBooster’s settings:

Noise Suppression Strength: start at Medium. If you can still hear fan noise, move to High. If your voice starts sounding slightly processed or breathy, drop back to Medium or Low.
Voice Activity Detection (VAD): optionally pair this with noise suppression — it silences the output entirely during pauses, which eliminates any residual noise floor between sentences.

The virtual mic VoxBooster creates appears in Windows sound settings as a standard audio device. Discord, OBS and every other application see it as a normal microphone.

Specific Noise Types and How to Handle Them

Fan and air conditioning hum

Low-frequency steady-state noise. RNNoise handles this well. If you also have a significant 50/60 Hz component from electrical interference, a narrow notch EQ filter at that frequency before the noise suppressor can help — it is a static problem that EQ solves more efficiently than ML inference.

Mechanical keyboard clatter

Intermittent broadband transients that arrive in tight clusters at typing speed. This is where RNNoise noticeably underperforms versus AI denoisers. The transients are brief enough that they overlap with spoken consonant sounds (specifically stop consonants like P, T, K), making discrimination harder.

If you use a mechanical keyboard, physical mitigation helps: desk mat under the keyboard, O-rings on switches, or a directional microphone positioned to minimise pickup of the desk surface. Combined with an AI denoiser, keyboard noise becomes a non-issue.

Room echo and reverb

Echo is reflected sound that arrives at the mic a measurable time after the original. Reverb is the accumulation of many overlapping reflections that create a diffuse “room sound.” Neither is addressed well by RNNoise.

Physical treatment (acoustic foam panels, a reflection filter behind the mic, moving the mic closer to your mouth and reducing gain accordingly) is the most effective solution. Software can reduce echo, but it cannot fully compensate for a room with long reverb tails. For home setups, a reflection filter behind the mic and a cardioid polar pattern to reject room sound from the rear makes a bigger difference than any software setting.

Background speech

AI denoisers are trained to identify human voice frequency patterns. When the background noise is also a human voice (family member talking, TV in the background), the classifier has a harder job. Good AI implementations handle this reasonably by looking at spatial and spectral cues, but no software perfectly separates two simultaneous voices at a single microphone. Physical isolation — closing the door, moving to a quieter space — remains the best solution.

Mouse clicks

Similar to keyboard clatter but with a different transient profile. Generally easier to suppress than keyboard noise because the click rate is lower and the transients are less sustained. Most AI denoisers handle mouse clicks cleanly at medium suppression settings.

Latency: What the Numbers Actually Mean

A common concern is whether noise suppression adds perceptible delay to your own voice. Here is how to think about it:

Below 20 ms: completely imperceptible during live speaking. You will never notice.
20–50 ms: technically present but not audible as echo during normal speech. Edge case: if you monitor your own mic through headphones, some people notice it in this range.
50–100 ms: begins to feel like a slight lag if you monitor your own processed voice in real time.
Above 100 ms: noticeable echo when monitoring. Unacceptable for live performance use.

RNNoise adds approximately 10 ms. Real-time AI denoisers target 15–30 ms. Neither is an issue for the vast majority of use cases. The only scenario where noise suppression latency matters is if you are doing professional music recording with real-time monitor mix — which has its own low-latency requirements that apply long before noise suppression enters the picture.

Noise Suppression and Voice Changers: Making Them Work Together

If you use a real-time voice changer alongside noise suppression, the ordering rule from earlier becomes critical. The common failure mode is:

User enables noise suppression in OBS.
User also enables a voice changer plugin in their DAW or a separate application.
The DAW’s voice changer processes the raw mic first, then OBS’s noise suppression runs on the changed voice.
The noise suppressor, encountering unfamiliar spectral content from the pitch-shifted voice, removes some of it along with the noise.

If you are running both features through separate apps, make sure the clean mic signal goes to the noise suppressor first, and only then to the voice changer. VoxBooster sidesteps this entirely because both features are integrated in one application with a fixed correct order.

For streamers running OBS virtual camera and voice effects simultaneously, check out the guide on how to use a voice changer on Discord for the full routing setup.

Does Discord’s Built-In Noise Suppression Make Dedicated Software Redundant?

Discord provides three levels of noise suppression (None, Low, High) and an echo cancellation toggle in Voice & Video settings. These use a combination of WebRTC echo cancellation and an ML-based denoiser applied to your incoming signal in the client.

The Discord suppression is fine for casual calls. Its limitations:

It applies only to the Discord application — no benefit in OBS, games or any other app.
You cannot control where it sits relative to other processing.
It only shows effect in Discord’s own output — your stream recording through OBS does not benefit.
At “High” level, some voice tones (particularly higher female voices) sound noticeably processed.

Turn Discord’s noise suppression off when using dedicated software to avoid double-processing, which degrades voice quality. Discord’s echo cancellation can remain on independently if you want it.

Evaluating Noise Suppression Quality

How do you actually judge whether your setup is working? Record a 30-second clip with your mic at the desk, speaking naturally while your noise sources run (PC fans, AC, keyboard). Then mute yourself mid-sentence for 5 seconds to capture the noise floor.

Check:

Noise floor (the 5-second silent section) — ideally below -50 dBFS. Below -60 dBFS is excellent.
Voice clarity — do your sibilants (S sounds) still sound natural, or slightly hissy/processed?
Consonant preservation — say words with P, T, K sounds. Are they crisp or slightly muffled?
Transition artefacts — right as you start speaking after silence, does the voice pop in cleanly or are there brief clicks?

If you pass all four checks, your noise suppression is well tuned. Adjust strength lower if you hear voice degradation; adjust higher if the noise floor test is above -40 dBFS.

For more on audio noise suppression fundamentals, the Wikipedia article on noise reduction covers the signal processing background in readable depth. The RNNoise project page has interactive audio demos comparing filtered and unfiltered recordings.

Combining Noise Suppression with Other Audio Features

Noise suppression pairs well with these additional processing stages:

Noise gate — place it after the suppressor as a second line of defence. The gate silences any residual noise floor during pauses that the suppressor reduces but does not eliminate. Set the gate threshold just above the suppressed noise floor level.
Compression — after suppression and gating, gentle compression (3:1 ratio, slow attack) evens out your vocal dynamics without re-amplifying noise. Because the noise is already removed, compression works cleanly.
Voice effects — last in the chain. Pitch shift, resonance, and character effects all produce better results on a clean, suppressed signal.

If you are also using noise suppression alongside a soundboard, note that soundboard audio usually bypasses the mic chain entirely — it routes directly to the virtual output. This means noise suppression does not affect soundboard sounds, which is the correct behaviour.

Frequently Asked Questions

How do I remove background noise from my microphone?

The most effective route is dedicated noise suppression software that processes your mic signal in real time before it reaches Discord, OBS or any other app. Tools like VoxBooster apply AI-based filtering at the system level, so every application benefits automatically without per-app configuration.

What is the best noise suppression software for streaming?

It depends on your hardware and workflow. Streamers who want a single app handling noise suppression, voice effects and soundboard in one go tend to prefer VoxBooster. Those already in the OBS ecosystem often start with the built-in RNNoise filter before deciding whether to upgrade to a dedicated tool.

Does noise suppression software add latency?

Yes, but modern implementations keep it very low. RNNoise adds roughly 10 ms. More powerful AI/ML filters add 10–30 ms depending on chunk size and model complexity. For live voice this is imperceptible. Heavier denoise models designed for post-production can add 100 ms or more and are not suitable for real-time use.

What is the difference between noise suppression and noise cancellation?

Noise suppression filters unwanted sound from a microphone signal in software — it analyses the audio and attenuates non-voice frequencies. Noise cancellation (like in ANC headphones) uses a physical microphone and speaker to produce an inverse sound wave that cancels noise acoustically. Both aim at the same result through very different means.

Does noise suppression work with any microphone?

Yes. Software-based noise suppression sits in the audio processing chain after the ADC stage, so it works with USB mics, XLR interfaces, laptop built-in mics and headset mics alike. A better microphone gives the algorithm cleaner input, but even a noisy laptop mic sees significant improvement.

Will noise suppression affect my voice quality?

Light-to-moderate noise suppression has minimal impact on voice. Aggressive settings or low-quality algorithms can cause voice to sound watery, robotic or clipped. The key is tuning suppression strength to your environment — enough to kill the noise floor, not so much that the algorithm starts eating consonants.

Can I use noise suppression and a voice changer at the same time?

Yes, and the order matters. Noise suppression should run first, before the voice changer processes the signal. If you layer them the wrong way around, the voice changer creates new harmonic artefacts that the suppressor then tries to remove, degrading quality. VoxBooster handles this ordering internally so you do not have to manage it manually.

Conclusion

Removing background noise from your microphone is not a single-setting fix — it requires understanding where noise suppression sits in your chain, which technology handles which noise type, and how to tune the strength without degrading voice quality. The free path (RNNoise in OBS) covers a lot of ground and is worth setting up even if you plan to move to dedicated software later. For system-wide coverage, keyboard and echo handling, and integration with voice effects, a dedicated tool earns its keep quickly.

VoxBooster includes AI-based noise suppression as a built-in first-stage processing step, meaning it cleans your mic before voice effects, soundboard routing, or anything else touches the signal. Whether you are streaming, in Discord calls, or on a work video meeting, it works across every app simultaneously through a single virtual microphone device. Check the pricing page for plan options.

Download VoxBooster — free 3-day trial, no credit card required.