Speech enhancement

Traditional noise reduction worked by subtracting a noise profile, which left the watery artefacts familiar from older podcasts. Learned models take a different route: they reconstruct the speech signal itself, which is why they can strip an air conditioner or a busy street from a recording and leave dialogue that sounds close to studio quality.

For creators this decides whether footage is publishable. Viewers tolerate imperfect video far longer than bad audio, and short form is often shot in rooms with no treatment, on phones, or over a call. Cleaning dialogue is usually the highest impact single change to a clip cut from a webinar, an interview, or an outdoor recording.

The nuance is that enhancement is restoration, not replacement. It cannot recover words lost to clipping, undo heavy compression from a video call, separate two people talking over each other, or add detail a cheap microphone never captured. Pushed too hard it also thins the voice and introduces a processed, underwater quality that is worse than the original noise.

In practice the order matters: enhance each speech source first, then balance levels between speakers, then add music and effects. Cleaning after a mix has been built means the music gets processed along with the voice, which produces pumping and smeared transients that are far harder to remove than the original noise.