Why noise gates ruin podcast audio

What a gate does

A gate watches the level. Above a threshold, audio passes. Below it, audio is silenced. That is the whole mechanism, and it was a reasonable answer in an era when there was nothing better.

The consequence is that the noise is still there whenever anybody is speaking — it is simply masked by the voice. It disappears in the pauses and returns the instant a word starts.

Why listeners notice without knowing why

That appearing-and-disappearing room tone is what people mean when they call a recording "pumpy" or say it sounds processed. The ear is extremely good at noticing a background that switches on and off, because nothing in the physical world behaves that way.

It also makes speech sound closer to the microphone than it is, then suddenly further away, which is fatiguing over an hour even when no single moment sounds wrong.

The other failure: clipped words

A threshold cannot distinguish a breath from the start of a softly spoken word, so a gate tight enough to catch the breathing also eats consonants. A gate loose enough to keep the consonants does not catch much of anything.

There is no setting that solves both, because the tool is answering the wrong question — how loud is this? — instead of the right one: what is this?

What to use instead

A model that separates speech from everything else removes the floor rather than hiding it, so the pauses and the speech are both clean and the background does not move. That is what a full-band neural de-noiser does, and it is what this application runs.

Then, separately, use a speech detector for the silence trimming, so what gets tightened is decided by whether anybody is talking rather than by a level. Two different questions, two different tools.

Fix the room first, though

None of this beats not recording the noise. Move away from the fan, put something soft on the hard surface behind you, and get the microphone closer to your mouth than it currently is.

Cleanup tools have become good enough that people skip these steps. They still produce a better result than any amount of processing does.

In short: A gate hides noise and clips words. Use speech separation for the noise and a speech detector for the silence.