How to clean up a talking-head recording

You recorded yourself talking for forty minutes. The content is there; the delivery is full of pauses, restarts and "um"s, the room has a hum, and the music you dropped underneath is fighting your voice.

All of that is mechanical, and the passes have a correct order. Doing them in the wrong order costs tokens and, in one case, makes you redo work.

Pass 1 — Trim before anything else

Cut the top and tail, and delete the takes you already know are dead. Nothing later in this list should ever run over material you were going to throw away, because two of the later steps are priced by length and all of them are faster on less footage.

Pass 2 — Cut the silences (free)

Select the clip and run Remove silences from its Audio tab. It scans for quiet stretches and proposes a cut for each one, with its time and length; untick any pause you want to keep, and one click executes the rest. Cuts ripple across every lane, so linked audio and music stay in sync, and a single undo restores all of them.

Do not accept every proposal. A talk with every pause removed is exhausting to watch — breath is how an audience keeps up. Keep the beat before an important sentence and the one after a joke.

This pass costs nothing, which is also why it belongs before the next one.

Pass 3 — Cut the filler words (billed by length)

Remove filler words transcribes the speech word by word and lists every "um" and "uh" with its timestamp, on the same untick-what-you-want-to-keep checklist. Each cut is given a little breathing room, so words never get clipped.

The scan is billed on the clip's length, and that is the entire reason for this ordering: run it after the silence pass and you are paying for a shorter clip. Same result, less money.

The lexicon is deliberately conservative — pure vocal sounds only — so it will not offer to cut real words like "like". And sometimes an "um" is characterful. Untick it.

Pass 4 — Repair the audio, cheapest option first

There is a ladder here, and most recordings are fixed on the first rung:

1. Export-time cleanup. Loudness normalisation and noise reduction are per-clip switches that cost nothing and handle mild hum and hiss. Try them before anything else.
2. AI denoise, for a recording that is genuinely rough. In the AI tab, switch the output to Sound and pick Rework a track — it takes your audio as its input instead of a prompt.
3. Voice isolation, when there is traffic, a fan or a conversation behind you. It lifts the voice off the background entirely, which is a bigger hammer and sounds like one; use it when the alternative is unlistenable.

If the picture is fine but the sound is unusable, note that a video clip's embedded audio can be split onto its own lane and edited — or replaced — without touching the picture.

Pass 5 — Levels and ducking

Set the voice as your reference, then bring the music in around 12 dB below it. Rather than riding that by hand, flag the music clip as Duck in its Audio tab: it dips whenever the voice lanes carry speech and recovers in the pauses, identically in preview and export.

Per-clip fade handles cover entrances and exits; overlap two fades for a crossfade. If you need something custom, volume keyframes are there.

One warning worth repeating: check the mix on phone speakers, not headphones. Music that sits politely under a voice on monitors routinely swallows it on a phone, and a phone is where this will be watched.

Pass 6 — Cover the jump cuts, if they bother you

Two hundred cuts in twelve minutes is the modern house style and mostly passes unnoticed. If a particular one jars — a big head-position jump mid-sentence — cover it: a b-roll insert or PiP over the join, a title card, or a short zoom change so the cut reads as a change of framing rather than a glitch.

Pass 7 — Caption last

Captions come last for one reason: they are generated against the audio as it exists, and every cut you make afterwards puts them out of step. Recutting after captioning means re-running the captions.

So lock the edit, then caption. Keep lines short — auto-captions are the part of this whole process that most affects how many people watch to the end, because most of them are watching muted.

The order that works

1. Trim the dead material by hand.
2. Remove silences — free, and it shortens what you pay for next.
3. Remove filler words — billed by length, so it runs on the tightened clip.
4. Repair the audio: free export cleanup, then denoise, then voice isolation.
5. Set levels and duck the music, then check it on phone speakers.
6. Cover only the cuts that jar.
7. Caption last, once the edit is locked.