A practical post-production workflow for cleaning up noisy video audio — and why prevention always beats the best plugin.
Background noise is one of the fastest ways to lose a viewer's trust. A shaky camera can feel cinematic; a room full of hiss, hum, or reverb just sounds unprofessional. The good news is that modern noise reduction tools — including AI-powered voice isolation — can rescue a surprising amount of bad audio. The bad news is they all have a ceiling, and crossing it means trading background noise for artifacts that sound worse. This guide walks through the full workflow: what to fix in post, how to fix it, and how to reduce the problem at the source so post-production stays manageable.
Put on headphones and listen to your raw audio in isolation. Identify whether the problem is broadband hiss, electrical hum, room reverb, low-frequency rumble, or intermittent ambient sound. Write it down. Each type requires a different tool and a different setting, and mixing them up wastes time and degrades the voice.
Set a high-pass filter at 80 to 120 Hz and roll off everything below that point. This removes handling noise, foot traffic through the floor, HVAC rumble, and the low-frequency content of electrical hum. This is a near-universal first step for voice audio and costs nothing in voice quality since the human voice carries almost no useful content below 80 Hz.
If the recording sounds like the speaker is in an echoey room, apply de-reverb before any other noise reduction. Set the plugin to analyze the room response and reduce the reverb tail — start at a moderate setting and increase until the voice sounds clear but not dry and lifeless. Apply too much and you will get artifacts; the goal is intelligibility, not a dead studio sound.
If you hear a steady tonal hum, use a de-hum plugin set to the appropriate frequency for your region — 60 Hz for North America, 50 Hz for most other countries. The plugin should also remove harmonics at multiples of the fundamental. Avoid using broadband noise reduction for hum; it will damage the voice and leave the hum largely intact.
For hiss, background noise, and general room ambience, apply a noise reduction plugin. If your software supports it, use a noise profile sample — typically a section of silence at the start of the recording — to teach the plugin what to remove. Start conservatively and increase the reduction only until the noise becomes inaudible. Alternatively, run the audio through an AI voice isolation tool, which handles noisy sources better at the cost of slightly homogenizing the voice character.
Once the noise floor is addressed, apply gentle compression to even out the dynamic range — a ratio of 2:1 to 3:1 with a medium attack and release works well for speech. Then bring the overall level up to match your delivery target using your editing software's loudness normalization or a dedicated loudness plugin. Check the result on multiple playback devices before exporting.
Bypass your entire processing chain and listen to the raw audio, then re-engage it. Make sure the voice still sounds natural and unprocessed — you should hear the noise disappear, not the voice change character. If the processed version sounds metallic, swirly, or over-compressed, dial back the most aggressive step and repeat until the balance is right.
Background noise is not one problem — it is a family of different problems that each require a different tool. Broadband noise is the constant hiss you get from a low-quality microphone, a high ISO camera mic, or a loud environment like an open-plan office or outdoor shoot. Hum is a steady tonal frequency — most commonly 50 Hz or 60 Hz electrical interference from lighting, power supplies, or a ground loop in the audio chain. Reverb and room echo occur when a microphone picks up sound bouncing off hard surfaces, making the speaker sound like they are in a bathroom or a gym. Each of these is addressed with a different processing tool, and stacking the wrong tools in the wrong order will make things worse, not better.
Before you reach for any plugin, listen critically to your raw audio on headphones. Identify which type of noise you are dealing with. Is it constant or intermittent? Is it broadband or tonal? Does the voice sound distant and washy, or is it bright but buried under hiss? That diagnosis determines your processing chain and sets realistic expectations for how clean the final result can actually be.
Every noise reduction tool works by separating the voice from everything else in the audio signal. The cleaner the original recording, the more accurately the tool can make that separation. When the noise floor is low relative to the voice — meaning the speaker was close to a decent microphone — noise reduction works well and artifacts are minimal. When the noise floor is close to or louder than the voice, the tool has to make aggressive guesses, and those guesses introduce audible problems: a metallic or watery quality on the voice, swirling artifacts during quiet passages, and a loss of the natural air and texture that makes a voice sound real.
The industry shorthand is that you can typically remove noise that is 20 to 30 decibels below the voice level without serious artifacts. Anything noisier than that, and you are in triage territory — you might salvage the audio enough to be usable, but it will not sound great. No plugin, including current AI tools, can fully reconstruct audio information that was never captured. Knowing this ceiling prevents you from spending hours on a clip that was always going to sound marginal and from overselling results to a client.
A newer category of tool uses machine learning to identify and extract the human voice as a separate element, essentially treating noise removal as a voice detection problem rather than a subtraction problem. These tools — available as standalone apps, plugins, and cloud-based processors — are trained on large datasets of speech and have become significantly better at preserving voice naturalness than traditional spectral noise reduction. They work particularly well on broadband noise and can handle noisier source recordings than older methods.
The tradeoff is that AI voice isolation tools can occasionally remove non-voice audio you want to keep — room ambience, music beds, or a guest who speaks quietly and falls below the model's confidence threshold. They also tend to homogenize the voice slightly, smoothing out the acoustic character of the recording environment. For dialogue, interview, and podcast content this is usually fine. For music performance or ASMR content where the acoustic texture is intentional, use AI tools sparingly or not at all. Run the processed and unprocessed versions side by side before committing.
Electrical hum requires targeted tonal removal rather than broadband noise reduction. A de-hum plugin removes the fundamental frequency — 50 Hz in most of Europe, the Middle East, and Asia; 60 Hz in North America — plus its harmonic series at multiples of that frequency. Applying broadband noise reduction to a hum problem will make the voice sound unnatural while barely touching the hum itself; use the right tool for the job.
De-reverb is one of the harder audio restoration problems. It works by estimating the room impulse response — how long reflections take to decay — and attenuating those tails. Results vary widely depending on room size, the amount of reverb relative to the direct signal, and which tool you use. A lightly reverberant room recorded with a reasonable microphone often responds well. A large, echo-heavy space recorded with a camera-mounted mic is much harder to fix. Apply de-reverb before noise reduction in your processing chain, since reverb removal can expose underlying noise that should then be addressed separately.
After noise reduction and tonal cleanup, level your audio for your delivery platform. Dialogue for YouTube and social video should sit between -14 and -16 LUFS integrated loudness. Use a loudness meter rather than eyeballing the waveform, and apply a limiter at -1 dB true peak to prevent distortion on playback devices. Compression before the limiter can help even out peaks from close-talking or emphatic speech, but keep it subtle — over-compressed dialogue sounds pumped and tiring to listen to.
Every minute spent on audio cleanup in post is a minute you pay for preventable problems. The single most effective noise reduction tool is a close-placement directional microphone. A cardioid or hypercardioid microphone placed six to twelve inches from the speaker's mouth rejects far more room noise than any plugin, because distance and directionality change the signal-to-noise ratio before the signal is ever recorded. A basic dynamic microphone used correctly will outperform an expensive condenser used badly.
Room treatment matters more than most video creators realize. You do not need a professional acoustic treatment to make a meaningful difference — moving the shoot away from hard parallel walls, adding soft furnishings, or hanging a moving blanket behind the camera can reduce flutter echo and room resonance substantially. Filming outdoors introduces wind and ambient noise, so position your subject to minimize exposure, use a windscreen or dead cat on the microphone, and prefer a lavalier mic clipped close to the body over a camera-mounted mic on a windy day.
It depends on how loud the noise is relative to the voice. If the noise floor is well below the voice level, modern noise reduction and AI voice isolation tools can reduce it to the point where it is inaudible without noticeably affecting voice quality. If the noise is loud relative to the voice — for example, a street interview with heavy traffic close by — you can reduce it substantially but not eliminate it without introducing artifacts. There is no tool that can fully reconstruct audio information that was drowned out at the time of recording.
Traditional noise reduction works by analyzing the frequency profile of the noise and subtracting it from the overall signal. AI voice isolation takes a different approach: it uses a model trained on human speech to identify the voice as a separate element and output it independently of everything else. AI tools generally handle noisier recordings better and produce fewer metallic artifacts, but they can occasionally over-process a voice or remove quiet speakers. Both approaches have a place depending on the source material and how noisy the original recording is.
This is the most common artifact of over-aggressive noise reduction. It happens when the algorithm removes too much of the signal in its attempt to attenuate noise, and the remaining audio has gaps and inconsistencies that your ear perceives as a swirling, hollow, or metallic quality. The fix is to dial back the reduction amount — aim for the minimum reduction that makes the noise inaudible, not the maximum the tool allows. If you are already at a minimal setting and still hearing artifacts, the source recording may simply be too noisy to clean without compromise.
De-reverb is a processing tool that reduces the room echo baked into a recording. If your speaker sounds distant, hollow, or like they are in a large space, you need de-reverb rather than (or in addition to) standard noise reduction. Room reverb is a different problem from background noise — it is the acoustic signature of the space, not an external interference. Most video creators who film in untreated rooms with camera-mounted microphones deal with some level of reverb. Apply de-reverb before other noise processing for best results.
A better microphone placed correctly will always produce cleaner results than the best post-production tools applied to poor source audio. Noise reduction can reduce noise that is already quiet; it cannot manufacture audio quality that was never there. A directional microphone placed close to the speaker changes the physics of the recording — it captures more voice and less room — in a way no plugin can replicate. If budget is limited, a close-placement dynamic microphone and a quiet room will deliver better results than an expensive condenser in an untreated space with heavy post-processing.
A full-episode workflow that keeps the conversation natural, your brand consistent, and your clip pipeline full.
A practical, step-by-step workflow for turning raw interview footage into a tight, story-driven video — from paper edit to final export.
Studio432 edits concert footage, music videos, gaming content, vlogs and short-form for creators worldwide. Send the footage, get a quote by email.
Start a project →