How to Overwrite a Word in Descript

Most “overwrite a word in Descript” posts collapse three jobs into one button. Correct fixes the page and leaves the waveform. Delete removes the token and the time it occupied. Regenerate — the current public name for Overdub — types a new word into an authorized clone and asks you to preview the join. This guide is that last path: one wrong token, one generate, one listen — using the one editor we send for this brief: Descript.

When a one-word overwrite is the job — and when you should Correct or re-record instead

Use this pass when the rest of the take is already the episode and one spoken token is wrong. A year you flipped, a guest’s surname you clipped, a product that shipped under the old name: you do not want a new booth hour. You want that one word said correctly in a voice that still sounds like the person who sat in that room. Official change-what-was-said copy names that job — “a wrong spoken word in the original audio.” Regenerate is the published control. A custom AI Speaker is the published voice. Consent is the published gate.

Skip this pass if the audio already said the right word. That is Correct. Official Correct your transcript help: double-click the word, click Correct or press C, type the spelling, then Correct this instance or Correct All. The media does not move. Captions and exported text follow the page. Pressing C on a misspoken date will make the script look honest and leave the mouth saying last year. We name that trap so you do not ship it.

Skip the clone if you can say the word again in ten minutes on the same mic. A real pickup will match the room without a consent script. Skip this pass if the remaining job is a sentence you do not want at all. Deleting the tangent at 18:40 is a highlight-and-delete job on How to cut on the transcript in Descript — an ordinary guide link, not a second money hop. Skip both if the file still sounds like a kitchen: How to use Studio Sound in Descript. Type on the picture is How to add captions in Descript. The file out is How to export a video from Descript. A whole misspoken line is How to overdub a line in Descript. Those pages hop to Descript too. This one already will.

The seven-step Descript one-word replace

1. 01

   Confirm you own the voice — and that the problem is one spoken word, not a spelling on the page

   This pass is for a host word you said, or a voice the speaker has authorized you to clone. Official create-AI-Speaker help will not authorize a deceased person, a non-consenting speaker, someone who cannot record the training statement, or audio that already came from an AI. A guest surname you flubbed is not automatically yours to generate. If you do not own the voice, stop and re-record. Then be honest about the job: if the audio already says the right name and the transcript spelled it wrong, that is Correct (C) — official Correct your transcript help changes the page and leaves the media alone. Stay here when the mouth said the wrong token.

2. 02

   Open the composition that already has the take — not a blank Write-mode script

   Regenerate (the current public name for the job people still search as Overdub) replaces words on a recording you already have. Official replace-what-was-said copy: use it when you misspoke, want to update a phrase, or need to fix a line without going back to the mic — including a wrong spoken word. Open the project and the composition that holds the cut. If the file is not in yet, import from your computer, from Zoom if that is the source, or from a phone upload. Wait for the script. Official add-a-file help: only script media is transcribed. A file sitting only as a layer is not a word you can overwrite. Do not start a new composition in Write mode and type a narrator. That is generate text-to-speech, a different brief.

3. 03

   Find the one wrong token in the script — a date, a name, a product — not the sentence around it

   Scroll to the misspeak. Highlight that word. Official Regenerate help says to highlight the section of script you want to change, then type the new word/s. For this page the section is one token: “2025” that should have been “2026,” a guest’s last name, a SKU, a city. Do not highlight the clause. Do not highlight the next scene. The line-length replace lives on the overdub-a-line how-to. Official limits still apply even to one word: each Regenerate action is capped at 250 characters; a selection that crosses a scene boundary or includes ignored text will not run; Regenerate needs original recorded audio on both sides of the selection. If the remaining problem is a three-minute ramble, that sentence still has to leave on the cut-on-transcript how-to. If the remaining problem is a fridge, that is Studio Sound, not a rewrite.

4. 04

   Assign the speaker, then train or pick an authorized custom AI Speaker

   Official change-what-was-said help: to replace what was said, the original speaker must have a custom AI Speaker created and authorized, and that speaker must be assigned on the script. Label the host first — Identify speakers after detection, or type @ / Change speaker. Official create path: Add speaker → Create speaker → name it → open the speaker’s … menu → Enable speech generation. That opens the consent and authorization script. Record it in English, on the same mic you used for the episode if you can. Speak naturally. Submit. Overview copy says a custom AI Speaker can train from as little as 30 seconds. If the host already authorized a clone on this Drive, pick that speaker — do not train a second one “just for one word.” Stock AI speakers are for writing a new voiceover. Official replace-what-was-said help does not treat a stock voice as a substitute for an authorized custom clone on the original speaker. Regenerate does not support stock voices for this job.

5. 05

   Open Regenerate, type the one correct word, generate the patch

   Official path: click Regenerate on the selection toolbar, or press D. A dialog opens with a text field. Type the word that should have been said — one token, not a rewritten paragraph. Then click Regenerate in the dialog. The script text pulses while it processes. Public 2026 help does not name a toolbar control called Speak for this job, and it does not name Overdub on the button — changelog copy is explicit that Overdub is now part of Regenerate. Official overview also names the neighbor job: Smooth or improve a section, where the script stays the same. Stay on change-what-was-said when the word itself is wrong. On a video episode with a single script track, the control defaults to Audio & video. Open the caret next to Regenerate and choose Audio only if you only want the word, not a mouth-move. Sequences do not support video regeneration. English audio and video only on the published Regenerate page.

6. 06

   Preview the new word in the sentence — Play, Retry, or Restore

   Official Regenerate toolbar after the pass: Play (preview the regenerated section), Restore (revert to the original recording), Retry (another variation), Accept (keep it). Turn Play on before you Accept. Listen for a clone that jumped rooms — a drier booth than the kitchen-mic take, a brighter S, a cadence that no longer sounds like the syllable before it. One wrong word is easier to hear than a rewritten sentence: the join is two phonemes wide. Official pricing copy says Regenerate is meant to match the surrounding tone. The published way to check that is preview, not a room-tone slider we can confirm. Official troubleshooting: if the clone is wrong, create a new AI Speaker with a fresh recording on the usual mic; there is no way to add more samples to an existing AI Speaker. Retry once or twice. If the third pass still sounds like a different room, Restore and re-record the word.

7. 07

   Accept, listen to the word before and after, then export if the episode is otherwise done

   Accept keeps the result. On video, official help clips the original picture to the regenerated section, moves it into the layer lane, and hides it. First-time video regeneration asks for a separate consent because it modifies a speaker’s visual performance. Then play the patched word in context — the breath before, the consonant after. Official AI-Speaker troubleshooting notes that generated speech often includes a little extra on either side for flexibility; convert the AI clip to an audio layer if you need to trim or crossfade like recorded audio. Official TTS help: timeline edits, fades, and precise playback want that conversion. Then Export → Local export if this was the last fix. That file-out pass is its own how-to. Published audio formats are .m4a, .wav, and .mp3; video local export is MP4. Listen to the file that left the app. If you still need captions, a transcript cut, or the room, those are ordinary site paths — this page already hopped to Descript.

Overwrite the spoken word vs Correct vs re-recording

“Overwrite a word in Descript” is a search. The decision is whether the mouth was wrong, or only the page. The Descript columns match published Regenerate, Correct-your-transcript, custom AI Speaker, and consent help as of August 2026. Re-recording stays qualitative here — we are not inventing a second hop for a booth you already own.

Correct vs Regenerate — write to the control that changes the mouth

Official selection-toolbar help puts both controls on the same hover strip. Correct “will allow you to correct the text of your script without affecting the audio.” Regenerate is “make a correction or smooth out an edit.” Those are not synonyms. If the guest’s name is spelled wrong on the page and said correctly on the tape, press C. If the guest’s name is spelled right and the host said the other sister, press D.

Descript’s changelog is the cleanest sentence we have for the search: Overdub is going away as a label, and replacing your script with AI Speech is now part of Regenerate. Highlight a section, click Regenerate on the script toolbar. If you edit the text, you need voice authorization for that speaker — the same gate Overdub used, under a different name. We are not going to keep teaching a retired button.

Public 2026 help does not name a toolbar control called Speak, Overwrite, or Patch Word for this job. Official change-what-was-said steps: highlight the section, click Regenerate or press D, type the new word/s, click Regenerate again. Official overview splits the neighbor job — Smooth or improve a section — where the script stays and the cut is ugly. Stay on change-what-was-said when the token itself is the problem. If a third-party write-up still says Overdub, the action we can confirm is Regenerate.

Train or pick the speaker — one word still needs a clone

Official getting-started copy still gives you two starts: record in Descript, or import an existing file. Import paths they publish include upload from your computer, import from Zoom, and upload from a phone. Put the file in the script. Wait. Media minutes start when the file is in the project. Do not re-import an hour you already radio-edited “just to overwrite one word.”

Label the person who misspoke. Automatic speaker detection can run on import; Identify speakers is the published follow-up. In the script, type @ or select text and Change speaker. Official Regenerate help will not rewrite the take until a speaker is assigned and that speaker has an authorized custom AI Speaker. A label that still says Speaker 1 is not a clone. A one-word highlight does not skip that assignment.

Official create path: Add speaker (usually at the top of the composition) → Create speaker → name it. Click the new label, find the name, open the … menu, choose Enable speech generation. That is the published action. It displays the consent and authorization script. Choose a microphone. Record. The statement must be read in English, even if you later generate non-English text-to-speech. Speak naturally, with varied tone — official copy says the delivery of that script shapes the clone. Review. Re-record if the take is noisy. Submit. Confirmation usually arrives within minutes. Overview copy: as little as 30 seconds of audio.

If that host already authorized a clone on this Drive, pick it. Assign that speaker. Do not train a second “better” one unless the first one failed a headphone pass — official troubleshooting is explicit that you cannot add more samples to an existing AI Speaker. A new speaker is a new recording, on the usual mic, in the tone you want (casual, formal). Stock AI speakers live on the same dropdown for Write-mode narration. Official replace-what-was-said help: Regenerate does not support stock voices for rewriting the original recording. We are not going to invent a “quick stock patch for one word” override.

Official list of voices they will not authorize: a deceased individual, a non-consenting speaker, someone unable to record the training and authorization statement, and audio originating from an AI or artificial source. Bypassing the process is a Terms of Service breach. We are describing their gate. We are not going to invent a “clone my guest from the ISO” button because the misspeak was only one syllable.

Highlight the token, type the correction, then preview the join

Highlight the one wrong word. Official run-through: Regenerate on the selection toolbar, or press D. A dialog opens with the current text. Type the word that should have been said. Click Regenerate. The script pulses while it works. Official overview copy is explicit that this family of tools can “replace a word or phrase.” We are staying on the word.

Official limits we will not sand off. Each action is 250 characters — a single token is nowhere near that, but the cap is still the published one if you accidentally highlight a paragraph. A selection that crosses a scene boundary or includes ignored text will not run. Regenerate needs original recorded audio surrounding the selection — if the word is bordered only by other AI speech, or has no surrounding audio, it will fail. That is why a one-word patch in the middle of a real take is the honest fit, and a one-word patch between two earlier regenerations is not. English only on the published Regenerate page.

On video with a single script track, the default is Audio & video; the caret next to Regenerate switches to Audio only. Sequences do not support video regeneration — the Audio & video option stays gray until you are on a single script track and have given video-regeneration consent. First-time video regen asks, because it modifies a speaker’s visual performance. Official help will also refuse B-roll, multi-speaker recordings, and shots where the face is not clearly visible. If you only needed the syllable, Audio only is the published narrower choice. We describe that caret. We do not invent a “lipsync this one word” toggle beyond those names.

Then listen on purpose. Official toolbar: Play, Restore, Retry, Accept. Play the new word in the sentence before and after it. A one-word join is unforgiving. A brighter S, a missing T, a clone that jumped rooms — you will hear it against the real syllable on either side. Official pricing copy: Regenerate is meant to match the surrounding tone and smooth awkward cuts. That is the claim. The published check is your ears. There is not a public slider we can confirm named Room Match, Presence, or “make it sound like the kitchen.” If the clone is brighter, drier, or faster than the take it sits in, Retry. If the third variation still jumped rooms, Restore. Train a new AI Speaker on the episode mic if the clone was built in a different space — artifacts in official troubleshooting usually came from the training audio (static, traffic, appliances, music, mouth noise).

Accept keeps the result. On video, the original picture for that section is clipped, moved to the layer lane, and hidden. Official AI-Speaker troubleshooting: generated speech often includes a little extra on either side so you can trim the join. Convert the AI clip to a standard audio layer when you need word-boundary edits, fades, or a crossfade — official TTS help says AI speech is not recorded audio until you convert it. Do not run Studio Sound on the generated clip until that conversion; official cleanup help says the effect will not apply to generated speech until you do. Think twice even then — you are re-processing a voice that was never in that room.

Underlord can be asked to fix a word. Treat that like a junior editor who will also burn credits. Watch the same preview. Full Underlord access is a plan question on the August 2026 grid; we do not invent a checkout total.

Export the patched file — or stop and re-record the word

If this was the last fix, leave with a file. That working path is How to export a video from Descript — Export in the upper-right, Local export, MP4 on video or .m4a / .wav / .mp3 on audio. Official help notes that .wav does not carry artwork or chapter markers. Descript does not mint your RSS URL. Upload the enclosure to the host that already serves the feed. We do not invent a host hop.

Official pricing copy on Regenerate: after you hit the specified limit, generated speech becomes “jibber” “jabber.” That is the published empty-tank sound, not a joke we added. Free-plan AI Speech, Regenerate, and custom voice clones are Limited on the August 2026 grid. Sit on a paid plan before the patched word is the one subscribers get. Confirm checkout on descript.com/pricing. We do not print a commission rate.

When the next job is captions, a cut, the room, or the file out

Stay here when one spoken word is wrong and you own the voice. Hand the composition to captions when the remaining job is type on the picture: How to add captions in Descript. Hand it to the transcript cut when a story still has to leave: How to cut on the transcript in Descript. Hand it to Studio Sound when the room is the problem: How to use Studio Sound in Descript. Hand it to export when the take is otherwise done: How to export a video from Descript. None of those are Try buttons on this page. The only hop here is Descript.

A whole misspoken sentence is still How to overdub a line in Descript. The um pass is How to remove filler words in Descript. Finding 5–15 verticals in a finished episode is a clipper job — start at Descript vs Klap if you have not picked that stack yet. None of those names are hops on this page.

We do not print an affiliate commission rate, and we do not invent a “this will earn you $X” line. The product question is whether the patched word still sounds like the person in that room after you previewed it — and whether you were honest about the voices you do not own.