5 Practical Ways to Clean Background Noise From Voice Recordings

Background noise is the most common reason a perfectly good recording sounds amateur. A ceiling fan, a laptop cooler, a fridge two rooms away, a neighbour's dog — none of it is loud, and all of it is there. The mistake beginners make is reaching for a single "denoise" button and dragging it to maximum, which trades the hiss for a watery, metallic voice that is somehow worse. Cleaning noise properly is a sequence of small, targeted moves, not one heroic setting. Here are five approaches, in the order an engineer would actually use them, with the numbers that make them work.

Before you touch a plugin, understand what you are fighting

Noise comes in two flavours, and they need different treatment. Broadband noise — hiss, fan whir, air conditioning — is spread evenly across the frequency spectrum. Tonal noise — a 50 Hz mains hum, a whining charger, a fluorescent buzz — sits at specific frequencies and stands out as a pitch.

Your ears and a spectrogram can tell them apart instantly. Broadband noise looks like a grey haze across the whole display; tonal noise looks like bright horizontal lines. This matters because a high-pass filter kills hum but does almost nothing for hiss, while a spectral denoiser handles hiss but can leave a hum ringing. Match the tool to the noise.

1. Start with mic technique and gain staging, not software

The cheapest noise reduction is the noise you never record. Every dB of noise you avoid at the source is a dB you do not have to remove later, and every dB you remove in post costs you a little bit of voice.

• Get closer. Halving the distance between mouth and mic raises the voice by roughly 6 dB while the room noise stays the same. That is a 6 dB improvement in signal-to-noise for free.
• Set gain so peaks hit around −12 to −6 dBFS. If you record too quietly and boost later, you amplify the noise floor along with the voice. Recording at a healthy level keeps the noise down relative to the signal.
• Kill the obvious sources. Turn off the fan, the AC, the fridge's ice maker. Record a five-second "room tone" clip with nobody speaking — you will need it later.

A room treated with a thick rug, curtains, and a bookshelf behind the mic will outperform a cheap plugin every time. Foam panels help with reflections; they do very little for a hum.

2. High-pass filter to remove rumble and hum

Almost every voice recording has energy below the range of the human voice that contributes nothing but mud. Male voices rarely go below 80 Hz; female voices rarely below 150 Hz. Everything beneath that — desk thumps, traffic, mains hum at 50 Hz — can go.

Apply a high-pass filter set to 80–100 Hz for male voices and 100–120 Hz for female voices. Use a gentle 12 dB/octave slope rather than a steep brick wall, which can introduce its own ringing. You will hear the recording get tighter and cleaner with no change to intelligibility.

If you have a persistent 50 Hz hum (India runs on 50 Hz mains; the US and parts of the Americas use 60 Hz), a high-pass alone may not be enough. Add a narrow notch filter at the hum frequency and, if it is still audible, at its harmonics — 100 Hz, 150 Hz, 200 Hz. Set the notch to a high Q (around 20–30) so you only remove the offending pitch and not the neighbouring voice content.

3. Use a noise gate to silence the gaps

A gate is a volume switch that mutes the signal when it drops below a level you set. It does not clean the noise during speech, but it makes the pauses perfectly quiet, which hugely improves the perceived quality of a recording.

Set the threshold just above the noise floor but below the quietest speech. If your room tone sits at −55 dBFS and your softest syllables peak around −35 dBFS, a threshold near −45 dBFS works. Add a hold time of 100–200 ms so the gate does not chop the ends of words, a release of 200–400 ms so it closes smoothly, and a range of −40 dB or so rather than a hard mute, which sounds less abrupt.

Be careful with gates on expressive speech. If a guest trails off softly, an aggressive gate will cut them off mid-sentence. It is often better to gate lightly and let a denoiser handle the rest.

4. Learn a noise profile with spectral reduction

Spectral noise reduction is the workhorse for broadband hiss. It works by analysing a sample of pure noise, building a fingerprint of it, and then subtracting that fingerprint from the whole recording.

The workflow is the same in every tool:

1. Select a clean sample of noise — your five-second room-tone clip, with no speech.
2. Ask the tool to "learn" or "capture" the profile.
3. Apply the reduction across the file, starting gently.
4. Reduce by 6–10 dB at first, not 20+. This is the single most important number. Most hiss becomes inaudible with 6–10 dB of reduction; pushing to 20 dB is what produces the underwater, robotic artefacts people complain about.
5. Increase in small steps and stop as soon as the hiss disappears. The moment you can hear the voice changing character, you have gone too far.

Listen for "musical noise" — a chirping, warbling texture left behind when the algorithm over-subtracts. If you hear it, back the reduction off.

5. Let an AI denoiser handle the hard cases

Traditional spectral tools treat noise as a static fingerprint, which fails when the noise changes — a passing car, a variable fan, a room with a mix of hum and hiss. Modern AI denoisers are trained on thousands of hours of paired clean and noisy speech, so they learn what a human voice should sound like and reconstruct it while suppressing everything else.

The practical advantage is that they need almost no setup. There is no profile to capture and no reduction amount to guess. The trade-off is that a very aggressive model can slightly dull consonants and remove the natural air of a voice, so choose a tool with a strength or "mix" control and start around 70–80 percent rather than 100. Compare the processed voice against the original on headphones before committing.

This is where a web tool genuinely helps. Offex Audio's free online cleaner removes background noise from voice recordings without any install, so you can upload a file, run it, and judge the result on your own audio in seconds.

Putting it together in the right order

The sequence matters as much as the settings. Do it in this order and each step makes the next one easier:

1. Clean up the source and re-record if the noise is severe.
2. High-pass filter at 80–120 Hz.
3. Notch any hum at 50 Hz and its harmonics.
4. Apply 6–10 dB of spectral or AI noise reduction.
5. Gate lightly to quiet the gaps.
6. Compare against the original and undo anything you cannot clearly hear as an improvement.

Conclusion

Noise cleaning is subtractive work, and every subtraction has a cost. The engineers whose recordings sound pristine are not using secret plugins; they are doing five small things well and refusing to overdo any of them. Fix the source, filter the rumble, notch the hum, reduce the hiss by a modest amount, and gate the gaps. Resist the urge to push a single control to maximum. Done in that order, with the numbers above as a starting point, even a noisy phone recording can come out sounding like it was captured in a proper studio.

Frequently asked questions

Try the free audio tools →

Related guides and tools