Cut the dead air out of a recording

Why this is harder than it sounds

Removing pauses is a one-line idea — find the quiet parts, delete them — and doing it that way produces a recording nobody wants to listen to. Three things go wrong, and each has to be handled deliberately.

Words get cut in half. Speech is full of brief quiet moments that are not pauses: the closure before a "p" or a "t", the gap between "back" and "to". A plain threshold sees those as silence. This detector uses hysteresis — it takes more level to decide speech has started than to decide it is still going — plus a hangover, meaning the quiet has to persist for the whole minimum gap before a cut is even considered.

Every join clicks. Butting two pieces of audio together leaves a step in the waveform, and a step is a broadband transient: a click. Each seam here gets an eight-millisecond equal-power crossfade, which is short enough to be inaudible as a fade and long enough to remove the discontinuity. The automated check on this tool asserts that the largest sample-to-sample jump in the output never exceeds the largest one in the input.

The result sounds rushed and inhuman. This is the big one, and it is why deleting pauses outright is the wrong goal. Listeners read gaps as punctuation — the beat before a punchline, the breath that starts a new sentence. Strip them all and the speaker sounds like they are being chased. So pauses here are shortened to a floor, never removed: a two-second stall becomes the 0.2 seconds you asked for, and the rhythm of speech survives.

The threshold comes from your recording, not from a number

Most tools ask you to set a silence threshold in dBFS, which assumes you know your own noise floor. Almost nobody does, and the right answer differs by thirty decibels between a treated booth and a laptop on a kitchen table.

Instead, this measures the file first. It takes the tenth percentile of short-frame energy across the whole recording — a robust estimate of the level the quiet parts actually sit at, which is not fooled by how much of the file is speech — and puts the threshold a fixed margin above that. The measured floor and the threshold used are both shown in the report, so you can see what it decided rather than guessing.

If it cuts into words, move sensitivity to cautious. If it misses obvious gaps, move it to aggressive: that usually means the room is noisier than the detector assumed.

Which settings for which job

• Interview or two-hander: cut pauses over 0.5 s, leave 0.2 s. Tightens the turn-taking without making people sound like they are interrupting each other.
• Solo narration or an audiobook: cut over 1 s, leave 0.35 s. Narration needs its pauses; you are only removing the stalls where the reader lost their place.
• Lecture or conference recording: cut over 2 s, leave 0.5 s. There is usually a lot of dead time while someone changes a slide, and very little worth tightening otherwise.
• A take full of retries: cut over 0.3 s, leave 0.1 s, then listen right through. This is the aggressive setting and the one most likely to need a second pass by hand.

Do this before you normalize, and before you transcribe

Order matters. Peak normalization is decided by the single loudest sample, so a chair creak in a gap you were about to delete can set the level for the whole file. Cut first, then set levels — and if the destination is a streaming platform, the LUFS normalizer is the one you want.

For transcription it matters for a different reason: Whisper and similar models sometimes invent text during long silences, a well-documented failure where a stray line appears out of nothing. Removing the dead air makes that less likely, and it is cheaper on any service that bills by the minute. Prep for Whisper afterwards.

What it will not do

It does not remove filler words. "Um" and "you know" are speech, with the same energy as everything around them, and finding them needs a language model rather than a level detector — which is what Descript and its competitors are doing when they offer that.

It does not fix a bad recording. If two people are talking over each other, or the room has a strong echo, the gaps are not clean and the detector has nothing reliable to work with. It also cannot tell a meaningful dramatic pause from an accidental one; if the timing is doing creative work, cut it by hand in the cutter.

And if all you need is to top and tail a file, trim silence at the edges does that and leaves the middle completely alone.