How to Remove Background Noise From Audio (What Actually Works)

Background noise is the most common reason a recording gets thrown away, and the most commonly misunderstood problem in audio. The hum, the hiss, the air conditioning, the traffic outside — people assume there is a button that removes it cleanly. There is not, and understanding why is what separates a recording you can rescue from one you cannot.

Here is the honest version: noise reduction is subtraction, not restoration. It removes the unwanted sound and part of the sound you wanted, because the two are mixed together in the same waveform. Used lightly it is close to invisible. Pushed hard it produces the watery, warbling, underwater artifact everyone recognises — and that artifact is usually worse than the noise it replaced.

A note on what this site does: AudioUtils converts, trims, compresses and joins audio in your browser. It does not do noise reduction — that needs spectral processing we do not offer, and pointing you at a tool that cannot do the job would waste your time. This guide covers the tools that can.

What "Background Noise" Actually Is

Different noise types need different treatment, and misdiagnosing it is why cleanup attempts fail.

Broadband noise is the constant hiss or rush that sits behind everything — preamp self-noise, air conditioning, room tone, distant traffic. It is spread across the whole frequency range at a steady level. This is the type noise reduction handles best, because it is predictable.

Hum is a narrow tone at a specific frequency, almost always mains electrical interference — 50 Hz in most of the world, 60 Hz in North America — plus its harmonics. Because it lives at known frequencies, a notch filter removes it surgically with very little damage to the rest.

Intermittent noise is a door, a cough, a chair scrape, a notification. It is not continuous, so noise reduction models cannot learn it. The fix is editing it out, not processing.

Overlapping noise is anything sharing the same frequency range as the voice — another person talking, music behind speech. This is the one nobody can fix properly. The information is genuinely blended; separating it means guessing, and guessing sounds like guessing.

How Noise Reduction Works

Nearly every noise reduction tool follows the same two-step logic, and knowing it explains every result you will get.

1. It learns a noise profile. You give it a stretch of audio containing only the noise — a second or two of "silence" before anyone speaks. It analyses the frequency content of that sample and builds a picture of what the noise looks like across the spectrum.

2. It subtracts that profile from the whole file. At every frequency, it reduces the signal by roughly the amount the noise profile says is noise.

The consequence follows directly: anything sharing frequencies with the noise gets reduced too. Voices contain energy across the same broad range as hiss, so aggressive reduction thins the voice, hollows it out, and leaves the characteristic warble as the algorithm makes moment-to-moment decisions about what to keep.

This is also why the noise profile matters so much. Profile a section that accidentally includes a word, and the tool will treat that word's frequencies as noise and carve them out of the entire recording.

Removing Noise in Audacity (Free, Step by Step)

Audacity is free, cross-platform, and remains the most practical option for this job.

1. Open your file in Audacity. 2. Find a noise-only section. Two seconds of room tone with no speech is ideal. If your recording has none, record ten seconds of silence in the same room with the same setup and paste it in — it is worth the trip back. 3. Select just that section. 4. Go to Effect → Noise Removal and Repair → Noise Reduction, and click Get Noise Profile. The window closes. That is expected. 5. Select the entire track (Ctrl+A / Cmd+A). 6. Open Noise Reduction again. Now set the three controls. 7. Preview, adjust, apply.

The three settings, and what they actually do:

• Noise reduction (dB) — how much to cut. Start around 6–12 dB. This is the control people ruin recordings with; the instinct is to push it to 30 and remove everything, and the result is unusable. If 12 dB is not enough, the recording may not be salvageable to the standard you want.
• Sensitivity — how eagerly it classifies something as noise. Higher removes more and damages more. Small adjustments only.
• Frequency smoothing — smears the reduction across neighbouring frequency bands, which softens artifacts. Around 3 is a reasonable starting point for speech.

Always preview before applying, and listen specifically for the warble on the ends of words. That is the first artifact to appear and the clearest sign you have gone too far.

Removing Hum Specifically

If the problem is a steady electrical buzz rather than hiss, do not use noise reduction. Use a notch filter at 50 Hz or 60 Hz depending on your region, and add notches at the harmonics (100/120, 150/180) if the buzz persists. In Audacity this is Effect → EQ and Filters → Notch Filter.

This is far gentler than broadband reduction because it removes a couple of narrow slices rather than attenuating everything. For voice work you can often add a high-pass filter around 80 Hz as well, which cuts rumble and handling noise below the range of the human voice at no cost.

The Rule That Beats Every Plugin

Fix it at the source. No processing recovers information that was never captured cleanly, and five minutes of setup beats an hour of repair.

• Get the microphone closer. This is the single highest-impact change available. Halving the distance roughly quadruples the signal power reaching the mic while the room noise stays where it is, so the ratio improves dramatically. A cheap microphone close to the mouth beats an expensive one across the room.
• Turn things off. Air conditioning, fans, fridges, dehumidifiers. Most constant hum in home recordings has a switch.
• Record a room tone sample. Ten seconds of silence at the start of every session. It costs nothing and it is what noise reduction needs later.
• Use a directional microphone pointed at the source, so it rejects sound arriving from elsewhere.
• Treat the room cheaply. Soft furnishings, curtains, a duvet behind the mic. Reflections are noise too.
• Watch your gain staging. Recording too quietly then amplifying afterwards raises the noise floor along with the voice. Get a healthy level going in.

When It Cannot Be Fixed

Being straight about this saves people hours:

• Speech buried under music or another voice. Same frequencies, genuinely mixed. Separation tools exist and are improving, but results on real recordings vary a lot and the output usually carries obvious artifacts.
• Clipping. If the input was too loud, the waveform tops were destroyed at recording. That is missing data, not noise, and nothing restores it.
• Heavy reverb. Not noise — it is your own voice arriving late off the walls. De-reverb tools help mildly and damage quickly.
• Wind on the capsule. Loud, low-frequency and overlapping the voice. A physical windshield prevents it; nothing removes it well.

If a recording is unrepeatable and this important, one honest option is to re-record the narration entirely and keep the original only as a reference.

For Transcription, Do Not Bother

A useful exception. If your only goal is a text transcript, heavy noise reduction usually makes accuracy worse, not better. Modern speech recognition is trained on messy real-world audio and copes with steady background noise well. What it copes with badly is the artifacting that aggressive cleanup introduces, because those artifacts do not resemble anything in its training data.

Feed transcription the cleanest unprocessed file you have. See best format for transcription for what to send.

A Sensible Order of Operations

1. Edit out intermittent noises — coughs, bumps, chair scrapes — by hand first. 2. Notch out any mains hum. 3. High-pass filter below the voice, around 80 Hz. 4. Apply gentle broadband noise reduction, 6–12 dB. 5. Normalise or compress last, so you are not amplifying noise you could have removed first. See what is audio normalization.

Do it in that order and each step has less to fight. Do noise reduction first and hardest, and everything after it works on damaged audio.

The Short Version

Noise reduction subtracts, it does not restore, so light-touch beats aggressive every time. Diagnose the noise type first — hum needs a notch filter, hiss needs broadband reduction, a cough needs the delete key, and overlapping speech needs a re-record. And the biggest improvement available to anyone is free: move the microphone closer before you hit record.