# Descript Overdub in 2026: Voice Clones, AI Speech and Regenerate

> How Descript Overdub works now: voice clones with English consent, text to speech in 21 languages, Regenerate word fixes up to 250 characters, and credits.

_Source: https://professionalstoolkit.com/articles/descript-overdub — The Professional's Toolkit · updated October 2026_

---

Descript Overdub still exists, just under new names: Descript now calls it Regenerate and AI Speech, and the voice clone behind it runs on ElevenLabs models, with v4 the current default since the October 2026 release. You need a paid plan for a custom clone, starting with Hobbyist at **$16 per person a month billed annually** or $24 billed monthly, and every generated line spends AI credits from a monthly pool. The real gate is consent, not money: the speaker must record a consent statement read in English before Descript will build the voice. Correcting a spoken word is English only and capped at 250 characters per action.


**250** — Characters per Regenerate action when changing spoken words, English audio only.



- **21** — Text-to-speech languages (help center)
- **English** — Required language of the clone consent statement
- **0.5 / 1.0** — AI credits per second of smoothed audio / video


## What Overdub is called now, and what each plan unlocks

Descript's March 2026 changelog describes Overdub as the legacy name, and timeline exports now label those tracks as Regenerate tracks. Two tools sit behind the old name. **AI Speech** turns typed text into voice, either a stock voice or your own clone. **Regenerate** rewrites or smooths what a real person already said, using that person's authorized clone. Help pages and older blog posts that still say Overdub describe the same feature family, so treat them as history rather than a separate product. The voice pickers draw on the ElevenLabs voice library, and Descript supports three of its models.

| Plan | Price per person | AI Speech access | AI credits |
|---|---|---|---|
| Free | $0 | Limited AI Speech trial | 100, one time |
| Hobbyist | $16 a month billed annually, $24 monthly | Custom voice clones, video Regenerate | 400 a month |
| Creator | $24 a month billed annually, $35 monthly | Same, plus lip sync and top-ups | 800 a month |
| Business | $50 a month billed annually, $65 monthly | Same, plus dubbing with proofread | 1,500 a month |

The table shows the pricing page as of October 2026 [1]. Hobbyist is the cheapest plan with a custom clone, and it allows one Editor seat. Credits reset each month with no rollover, and top-ups are sold only on Creator and Business. Dubbing with proofread is a Business feature; the help center says lower paid plans can also dub, so check the plan before buying for translation. Plan refunds are possible only within 48 hours of the invoice and Editor seats within 7 days, as our [Descript pricing guide](/articles/descript-pricing) explains.

## Step 1: record consent and build the voice clone

Start in the Speakers area of your Drive, or press @ in the script and choose to create a new voice clone. Descript then asks the person being cloned to record a consent statement, and that statement must be read in English even if you plan to generate speech in another language [2]. Clones are refused for a deceased person, a non-consenting person or audio that was itself AI-generated. The Terms forbid uploading training audio you are not authorized to use. After you submit the recording, the help center says the clone is typically ready within minutes.

Plan the clone before you record it. You cannot add more samples to an existing clone, and one clone cannot hold several speaking styles, so a calm narration voice and an energetic promo voice need two separate clones. The model is based on US English pronunciation. Accents may not survive. Descript suggests switching between ElevenLabs v2, v3 and v4 when a clone sounds off, which costs nothing to try except the credits for the new lines.


> 💡 **Make one clone per style:** A clone cannot take extra samples or hold several speaking styles, so record a separate consent and clone for each delivery you need.


A finished clone becomes a Drive Speaker that any member of the Drive can reuse across projects. Deleting it is permanent. Know what happens to the audio too: Descript's Terms let employees, vendors and Amazon Mechanical Turk workers listen to training and generated audio to check quality and misuse, and the privacy policy keeps voiceprints for up to three years after the last account access. Legacy clones without consent or valid training audio show yellow or gold text in the script and will not generate; the fix is a new clone with a fresh consent recording.

## Step 2: write text-to-speech lines and direct the delivery

Typing on a blank line of the Script editor switches on Write mode automatically, or you toggle it with Command+E on a Mac and Control+E on Windows. Put @ before a line to assign the speaker, stock or cloned, and Descript generates the audio as you type. Pulsing dark blue text means generation is running. An orange dotted underline means it failed, usually because the internet connection dropped. Text-to-speech covers 21 languages, while the pricing page lists native-sounding custom trained voices in 14.

Delivery is controlled with tone tags such as whisper, laugh, sigh and serious [3]. They work only on ElevenLabs v3 or v4, never on Multilingual v2, and custom tags are limited to 150 characters and still experimental. The old square-bracket prompts are gone: a script that still contains raw square brackets is blocked from generating until you remove them. Tone and voice can drift across long paragraphs, so generate in shorter blocks and listen to each one. Tags appear in exported transcripts but not in captions.

Stock voices are a moving target. They come from the ElevenLabs voice library, where community voices can be withdrawn at the owner's discretion, and older stock voices sit in a hidden Legacy category. If a project depends on one specific stock voice for months, a clone of a consenting narrator is the safer bet. When v4 arrives on your account, Descript switches the saved model to v4 once but does not regenerate audio you already made.

## Step 3: fix a spoken word with Regenerate

Regenerate covers the classic Overdub job: change a word without re-recording. Select the text, type the correction and Descript speaks it in the original voice. Three conditions apply [4]. The audio must be English, each action is limited to 250 characters, and the speaker needs a custom clone that the original speaker authorized. **Stock voices cannot change what a real person said.** That rule blocks the obvious misuse, and it also means a guest's line cannot be patched unless the guest recorded consent.

The same engine smooths awkward cuts. It works on English audio and video, and it refuses a selection that spans several edit points, a scene boundary, ignored text or a clip that was already regenerated. Each result can be played, retried for a new variation, kept or reverted. Hobbyist and up also get video Regenerate, which changes mouth movement to match the new words. It skips sequences, B-roll, multi-speaker recordings and shots where the face is not clearly visible. It is built for words and short phrases, not new clips. Two access rules catch teams late. Collaborators invited to a single project cannot use AI Speech at all, and text-to-speech and Regenerate fail over sequences: run them on a sequence and the video is removed. Do voice work on a single-track composition, then assemble the final cut.


> ⚠ **No patching without the speaker's consent:** Regenerate changes a real person's words only with a custom clone that person authorized; stock voices cannot be used for it.


## Step 4: convert to audio, fix pronunciation and timing

Generated speech stays live text until you convert it to audio. While it is live, you cannot adjust word boundaries, edit it on the timeline or add trims, fades and crossfades. Conversion unlocks all of that but is effectively permanent: text edits stop changing the sound, you cannot regenerate from the script, and tone tags turn into plain notes. Finish the wording first. Convert last.

Pronunciation follows the same order. Spell the problem word phonetically in the script, convert the clip to audio, then correct the spelling back in the transcript so captions read properly. To change the gap before or after a generated line, add a gap clip, which also requires the voice clip to be converted first. Convert to AI Speech, covered in the September 2026 changelog, works the other way and gives an existing recording a different voice. It handles a single video or audio track only and does not retime the video, so the new voiceover can leave gaps against the picture, which you cover with freeze frames, B-roll, a slowed clip or gap clips.


> 💡 **Finish words before converting:** Converting AI speech to audio is effectively permanent, so lock the wording and tone tags first, then convert to trim and fade.


## How far the credits stretch

AI Speech, Regenerate and dubbing all draw on AI credits, and Descript does not publish a flat rate for a line of text-to-speech. The one published voice rate is jump-cut smoothing: 0.5 credits per second of regenerated audio and 1.0 per second of video. By our arithmetic, a Hobbyist allowance of 400 credits covers about 13 minutes of smoothed audio or under 7 minutes of smoothed video, if nothing else used credits that month. Studio Sound, Underlord and avatars share the same pool, so real headroom is smaller.

The step up from Hobbyist to Creator costs $8 more per person a month billed annually ($24 minus $16) or $11 more billed monthly ($35 minus $24). That buys 400 extra credits a month and the right to buy top-ups. On annual billing the extra credits work out to $0.02 each, against about $0.10 per credit for the smallest top-up of 350 credits for $35. Hobbyist cannot top up at all, so a single seat that keeps running dry should upgrade rather than wait for the reset. Check the Usage tab after a test week before deciding. For the wider AI toolset, see our [Descript AI guide](/articles/descript-ai).


## FAQ

**Can I try voice cloning on the Free plan?**

Free gives only a limited AI Speech trial and 100 AI credits that do not renew; custom voice clones are listed from Hobbyist up. Descript offers a Free plan rather than a paid-plan trial.

**Where is a custom voice trained and hosted?**

Descript's security page says custom voices are trained and hosted on Google Cloud, and only from recordings of consenting speakers who read the designated consent script.

**Can Descript keep my voice recordings for research?**

The privacy policy says training data for AI Speakers may stay in research datasets after it is disassociated from your account. Check the Share data with Descript toggle under Settings > Profile.

**Can other AI tools rewrite a script in my cloned voice?**

Yes: Edit for Clarity has an optional Reword for flow step that generates AI speech, and it uses AI credits like the rest of AI Speech.


## The bottom line

Overdub lives on as Regenerate and AI Speech, and it is now a capable ElevenLabs-backed voice tool inside an editor rather than a separate add-on. Hobbyist at $16 per person a month billed annually is the entry point for a custom clone. The limits decide the fit: consent recorded in English, English-only word fixes of up to 250 characters, no sequences, and a credit pool shared with every other AI tool. Solo creators patching their own narration get the most from it. Teams that need to patch guests' voices without their consent cannot.



## Sources

[1] Descript Pricing - Plans for Every Creator, Free to Start Features Features Video editing Editing video in Descript is as easy as using docs — https://www.descript.com/pricing (2026-10)
[2] Descript Help: custom speaker — https://help.descript.com/ai-speech/custom-speaker (2026-10)
[3] Descript Help: tts — https://help.descript.com/ai-speech/tts (2026-10)
[4] Descript Help: change meaning — https://help.descript.com/script-editing/change-meaning (2026-10)
[5] Descript Help Center — https://help.descript.com/
[6] Descript Terms of Service — https://www.descript.com/terms
[7] Capterra, Descript reviews — https://www.capterra.com/p/230702/Descript/