How to remove background noise from video: fix it in order

To remove background noise from a video, work in this order: stop the noise at the recording, repair clipping first, apply the lightest noise reduction or voice isolation that makes the words clear, and set loudness last. Most bad results come from skipping the first step and overdoing the third, which leaves a voice that sounds robotic or underwater.

Free tools handle steady hum and hiss well: Audacity, CapCut and the free tier of Adobe Podcast Enhance Speech. Echo, wind and distortion are harder, and some of that damage cannot be reversed in software at all.

Before you record: remove the noise at the source

Every cleanup processor works on one ratio: how loud your voice is compared with everything else in the file. Moving the microphone closer improves it for free, without side effects.

• Mic distance. RØDE's rule of thumb for a lavalier is on the chest, about a hand-span (20-30 cm) from the mouth. A phone propped up across the room hears the room about as well as it hears you.
• Lav over phone mic. Most lavaliers are omnidirectional, so they forgive a sloppy angle; their real advantage is proximity. A built-in phone mic at arm's length is the setup most likely to need heavy repair later.
• Clothing. Keep the capsule clear of any fabric that can rub against it. Rustle sits in the same frequencies as speech, so noise reduction cannot remove it cleanly.
• Room. Switch off fans, air conditioning and the fridge for the take. Soft furnishings, curtains and carpet reduce echo; bare walls and tiled floors add it.
• Levels. Set input gain while saying your loudest line, laugh included. If the meter hits the top, those peaks are flattened for good.
• Room tone. Record a few seconds of silence before you speak. Profile-based noise reduction needs a sample of pure noise, and this is where it comes from.

Four fixes that sound alike and damage a voice differently

Editors blur these names together, and the wrong tool on the wrong problem turns a slightly noisy clip into an unlistenable one.

The popular tools and their real limits

Adobe's free tier and Audacity take audio only, so export the track as WAV, clean it and bring it back into the edit. Descript applies Studio Sound at the file level, so it changes every use of that file in the project. If you are cutting clips from a podcast recorded in an untreated room or over a call, try voice isolation before classic noise reduction, since the background there usually changes from minute to minute.

The free route: Audacity step by step

1. Export the audio from your edit as WAV and open it in Audacity.
2. Select a stretch that contains only noise, ideally the room tone you recorded. A longer profile works better.
3. Open Effect > Noise Reduction and click Get Noise Profile.
4. Select the whole track and open Noise Reduction again. The manual's advice is the lowest Noise Reduction (dB) value that brings the noise to an acceptable level; the default is 6.
5. Leave Frequency smoothing at the default of 6 bands or a little higher: per the manual, lower values favour music and higher values favour spoken word.
6. Switch to Residue and preview. If you can hear recognisable words in what is being removed, Noise Reduction or Sensitivity is too high. Switch back to Reduce and apply.
7. If two kinds of noise sit in different parts of the track, profile and reduce the first, then repeat for the second.

If random chirps appear after processing, that is the artefact the manual describes: Sensitivity was set too low, or the profile did not represent the noise elsewhere in the track. Take a new profile from a more typical section instead of pushing the reduction harder.

Loudness: what the platforms actually apply

Platforms publish very little about loudness for short video, so treat any single number as a reference point. The best documented case is YouTube. iZotope's platform table lists a -14 LUFS reference level, normalization always on, and no boost for quieter uploads. A voice left at -20 LUFS simply plays quieter than the videos around it, while a video mastered hotter than -14 is turned down and keeps whatever distortion the extra loudness caused.

For speech, the Audio Engineering Society's AES TD1008 recommends normalizing to -18 LUFS. Short vertical video competes in the feed with music-heavy posts, so around -14 LUFS integrated, with peaks kept clear of 0 dBFS, is a sensible delivery target. Measure it at the very end, after noise reduction and music, because both change the result.

Where Monty fits

Monty runs the cleanup pass on short vertical videos as part of the edit: denoise and declip, voice isolation, music ducked under the voice, and loudness mastered per platform before it publishes to YouTube Shorts, Instagram Reels, TikTok and Telegram. You record one take on your phone, and the audio chain runs together with the cuts, captions synced to speech and B-roll.

It cannot restore what the recording never captured. Hard clipping, wind blasts and a voice buried in a loud cafe stay damaged, so the checklist at the top still decides most of the result. Monty is also not a manual timeline editor: if you want to hand-tune de-reverb on one difficult clip, Audacity or Resolve is the right place. If you read your lines, the free teleprompter lets you keep the phone and its mic close while looking into the lens.