How to Remove Background Noise from Audio

"Noise" isn't one thing, and the tool that beats one kind is useless against another. Before you fix anything, identify what you're actually hearing:

• Broadband noise — steady hiss or 'shhh' spread across all frequencies: microphone self-noise, air conditioning, room tone, tape hiss. This is what classic noise reduction is built for.
• Hum and tonal noise — a low, constant drone at a specific pitch: 50/60 Hz electrical hum from bad grounding, a fan, a computer. Because it lives at one frequency, a narrow notch filter or a high-pass filter can remove it cleanly.
• Intermittent noise — sounds that come and go: a door slam, a cough, keyboard clicks, a car passing, someone talking in the background. This is the hardest kind, because there's no steady 'fingerprint' to subtract. Classic noise reduction can't touch it; you either edit it out by hand or use AI that understands what a voice is.

No software recovers a clean voice as well as capturing one in the first place. Every removal method trades away some quality; prevention costs nothing. If you can re-record, spend two minutes on this first:

• Get the mic close. Doubling the distance from the source roughly quarters its level relative to the room. A mic 15 cm from your mouth captures far less room noise than one across the desk.
• Kill the obvious sources. Turn off fans, air conditioning, and buzzing chargers. Close windows. Silence phone notifications. Move away from a humming fridge or computer.
• Tame the room. Hard bare rooms echo; soft furnishings absorb. Recording toward a bookshelf, a closet of clothes, or with a blanket behind the mic noticeably reduces reflections.
• Set levels right. Aim for peaks around -12 to -6 dBFS. Recording too quietly and boosting later also boosts the noise floor with it.

Before any dedicated noise reduction, try the simplest tool: a high-pass filter (also called a low-cut). It removes everything below a chosen frequency. The human speaking voice barely uses anything under about 80 Hz, but rumble, handling thumps, AC hum, and desk vibration live down there.

Set a high-pass filter at around 80 Hz for a typical adult voice (100 Hz for higher voices) and a surprising amount of low-end mud and hum just disappears, with zero damage to the words. For hum specifically, a notch filter at exactly 50 Hz or 60 Hz (whatever your country's mains frequency is) and its harmonics can surgically remove the drone. If EQ alone gets you there, stop — you haven't degraded the voice at all.

Audacity is free, open-source, and runs on Windows, macOS, and Linux. Its Noise Reduction effect is the classic answer to steady broadband hiss, and it works by learning a 'fingerprint' of your specific noise and subtracting it. The workflow trips people up because it has two distinct steps:

1. 1

   Select a noise-only sample

   Find a second or two where nobody speaks — just the hiss you want gone. Highlight only that region. This is the single most important step: the effect can only remove noise it has heard by itself.

2. 2

   Capture the noise profile

   Open Effect → Noise Reduction and click 'Get Noise Profile.' Nothing audible happens — you've just taught it what the noise sounds like. The dialog closes.

3. 3

   Apply to the whole clip

   Now select the entire track (Ctrl/Cmd+A), reopen Noise Reduction, and this time adjust the sliders and click OK. Start gentle: around 6–12 dB of reduction, sensitivity 6, frequency smoothing 3.

4. 4

   Listen for artifacts and back off

   Push reduction too hard and you get 'musical noise' — a watery, underwater warble, and a hollow voice. If you hear it, lower the reduction dB until the voice sounds natural again. A little remaining hiss beats a robotic voice.

Classic noise reduction is 'dumb' by design: it subtracts a spectral fingerprint and has no idea what a human voice is. AI voice isolation flips the approach. Instead of learning the noise, a neural network trained on millions of voice-and-noise examples learns what human speech sounds like, and keeps that while discarding everything else.

That difference matters most exactly where the classic tools break down:

• Intermittent, non-steady noise: background conversations, traffic, wind, clatter, echo — sounds with no fixed fingerprint to subtract.
• Noise that overlaps the voice in frequency: subtracting it with EQ or spectral methods would gut the voice too; a voice-aware model can separate them.
• 'One mangled recording, no clean sample': AI needs no noise-only region to learn from, so it works on clips where you never captured silence.

It isn't magic. Aggressive isolation can make a voice sound slightly processed or 'smaller,' and it occasionally clips a breath or a soft consonant. But for a noisy real-world recording that classic tools can't rescue, it's usually the difference between publishable and unusable.

FileMorf's Audio Studio pairs an in-browser editor with AI clean-up. The editing itself — trimming, cutting, arranging clips on the timeline — runs entirely in your browser and never uploads your audio. When you want the AI to isolate the voice, that specific step runs on secure servers, because voice-isolation models are far too heavy to run in a browser tab.

1. 1

   Open the Audio Studio and add your file

   Go to the AI Clean Voice panel and drop in your recording. You can trim to just the section you need first — local editing keeps that audio on your machine, so you only send what you actually want cleaned.

2. 2

   Run AI Clean Voice

   This uses AI voice isolation to keep the speech and strip broadband hiss, hum, room tone, and background sounds in one pass — including the intermittent noise that Audacity can't handle. It costs 2 credits per minute of audio.

3. 3

   Preview, then export

   Compare the cleaned result against the original. A free FileMorf account includes 6 AI credits per month — enough to clean a few minutes of audio at no cost. Files sent for processing are auto-deleted after the job; the local timeline never leaves your browser.