Descript: Edit Audio and Video by Deleting Words in a Text Document

You have just finished recording a 40-minute interview. Somewhere in the middle, you repeated the same sentence three times, and near the end you said "um" so often it sounds like a drumbeat. In a traditional editor, you would scrub through a waveform, hunt for the exact second, and slice it out. In Descript, you open the transcript, highlight the repeated sentence, and press Backspace. The audio is gone.

That simple gesture hides a lot of machinery. Understanding what happens between your keystroke and the silence in the audio helps you use the tool well — and notice when it needs your help.

How the text stays glued to the sound

When you import a file, Descript sends the audio to its servers and runs speech recognition on it. The result is a written transcript, but not a plain one: every word carries a timestamp pointing to the moment it was spoken. The transcript and the audio are two views of the same recording, kept in sync in both directions. If you drag the playhead in the audio, the matching word lights up in the text. If you delete a phrase in the text, the audio for that phrase is cut out.

This is why the editing feels immediate. You are not describing what you want removed; you are pointing at it in a form your eyes can scan faster than a waveform. The computer handles the frame-level cutting, including the small crossfades that keep the join from clicking.

The cleanup features that save the most time

Once the transcript exists, several chores become one-click jobs. Descript can find and remove filler words — the "um", "uh" and "you know" that clutter spontaneous speech — and it can shorten long silences. It also offers a studio sound option that reduces room echo and lifts a phone microphone closer to the quality of a proper recording. None of these are magic; they are pattern matches on the transcript plus audio processing, and they occasionally guess wrong.

The feature that surprises people most is voice regeneration. If you mispronounced a name or fluffed a line, you can type the corrected word into the transcript, and the tool synthesizes it using a model of the speaker's voice. That model is built from a sample of the person's speech and requires a recorded consent step before it can be created, which is a sensible guard against someone cloning a voice without permission.

Where the text-first approach shows its seams

Cutting words is easy. Cutting them invisibly is harder. If the speaker was moving their head while talking, removing several sentences in a row can produce a visible jump — the shoulders shift, the background flickers, the mouth lands in a different position. Descript includes smoothing options to soften those jumps, but they work best on small cuts, not on carving out a whole paragraph of a moving subject.

Transcription also has limits. Heavy background noise, strong accents, overlapping speakers and technical vocabulary all raise the error rate. A misheard word in the transcript is not just a typo: if you delete the wrong word, you delete the wrong sound. It is worth skimming the transcript before you start cutting, especially around names, numbers and anything a listener might quote back at you.

• Read the transcript first. Fix obvious misrecognitions before editing, so you cut what you meant to cut.
• Use smoothing on short cuts. For long removals in a moving shot, consider a cutaway image instead of a hard jump.
• Keep a copy of the original file. Text-based editing is fast, but it is still editing, and some decisions are hard to undo by ear.
• Check the consent step for voice models. If you plan to regenerate speech, the speaker needs to record that permission themselves.

The appeal of editing by text is not that it replaces judgment. It removes the mechanical part — finding the second, slicing the frame, nudging the crossfade — so you can spend your attention on the part that actually matters: whether the sentence should be there at all. The tool answers the how. The why is still yours.