Cleaning up speech with noise reduction and voice EQ

Most home and small-studio recordings carry some unwanted sound underneath the voice: the hum of a fridge, the hiss of a cheap microphone, the rumble of traffic outside. Video by Zaweba has two tools aimed squarely at speech: noise reduction, which takes the steady background noise down, and voice EQ presets, which shape the voice itself so it's easier to understand.

How the noise reduction works

The clever part is that you don't have to show it what the noise sounds like. For each clip, the noise reduction finds the quietest half-second, assumes that's the background noise on its own, and learns a noise profile from it. It then reduces sound matching that profile across the clip. Under the bonnet it uses FFmpeg's afftdn filter, a well established noise reducer.

Because it learns per clip, two clips recorded in different rooms each get their own profile. You don't need to find a clean gap and sample it by hand.

That does mean the quietest half-second matters. A recording that never pauses, with speech or music all the way through, gives the reducer less to learn from. If you're recording yourself, leave a second or two of silence at the start of each take. It costs nothing and gives the noise reduction a clean sample.

What it's good at, and what it isn't

Steady noise is the target: hiss, hum, fan noise, air conditioning, distant traffic that doesn't change much. These are the sounds a profile can describe.

Sudden noises are different. A door slam, a dog bark or a car horn isn't steady, so a profile won't catch them. For those, trimming, a quick volume dip with keyframes, or cutting around them works better. For electrical hum specifically, the sound effects library also has a hum remover.

Don't push it too hard either. Strong noise reduction can make a voice sound thin or watery. The goal is a background quiet enough that nobody notices it, not total silence.

Voice EQ presets

EQ changes the balance of low, middle and high frequencies. Voice EQ presets are set up for speech: taking away boominess, bringing forward the part of the voice that carries the words. Apply one to a dialogue clip and listen with and without. Often the difference is subtle but makes a voice sit more clearly over music.

If a preset gets you close but not quite there, the sound effect editors let you open an EQ and adjust its curve by dragging points.

Processed once, then cached

Both tools are processed once and the result is cached. The work is done up front, and after that playback is smooth because the app plays the cleaned version rather than recalculating it every time. If you change the settings, it processes again.

A sensible order

Clean first, then balance. Run noise reduction on the dialogue clips, add a voice EQ, and only then set the levels and the music underneath. Video by Zaweba also ducks music under speech automatically, and that works better on dialogue that's already clean, because the voice is clearer against the bed.

Listen on headphones as well as speakers. Leftover hiss that's invisible through laptop speakers can be obvious in earphones, and that's how much of your audience will hear the finished video.