Descript Transcription: Languages, Accuracy, Limits and Exports (2026)

Descript transcribes speech in 25+ languages automatically, and on current plans the transcription itself is not metered.^[1] What you pay for is media time: every upload and recording draws on a monthly pool of media minutes, whether or not you transcribe it. Free gives 1 hour a month, Hobbyist 10 hours, Creator 30 and Business 40 per Editor. Descript says accuracy can reach up to 95% with clear audio. Our verdict is that it is a strong transcription tool when you also edit the media, because in Descript the transcript is the editor itself. If you only need text, the media-minute cap is the number to watch.

What transcription costs you now

Every legacy plan was moved to two pools after 17 November 2025: media minutes and AI credits. Transcription draws on neither. Media minutes are spent the moment a file is uploaded or recorded, so a two-hour raw interview costs two hours even if you only keep ten minutes of it.^[2] Cleanup is a separate bill. Filler-word removal and most other AI tools spend credits, so the free part is the text, not the polish. Neither pool rolls over, and both reset at renewal; plan refunds run only 48 hours from the invoice date and Editor seats 7 days, as our Descript pricing guide explains alongside the other billing rules.

Some sources cost less than they look. A Rooms session counts its length once, not once per participant, so a one-hour call with five people uses 60 minutes. Re-using a file that already sits in the Media Library costs nothing, while uploading the same file a second time does. Cancelled uploads are free. When the pool runs dry, uploads stop until the reset or a top-up. Free also exports video at 720p with a watermark, which hurts a published video but not a text export.

Which plan fits your monthly volume

Up to 10 hours of new media a month, Hobbyist is the cheapest paid seat, though eligible nonprofits, students and educators get the same 10 hours on the Education / Nonprofit plan at $8 a month billed annually or $12 monthly. From 10 to 30 hours you need Creator. Media top-ups are one-time packs of 5 hours for $25, 20 hours for $80 or 50 hours for $150, usable for 12 months and sold only on the two upper self-serve plans. That is $5, $4 or $3 an hour. By our arithmetic, with the 50-hour pack, Creator plus top-ups beats a Business seat up to about 38 hours a month on annual billing and ties it at 40 on monthly billing. Business wins above that, or when you need its 1,500 AI credits and Brand Studio.

Getting media into the script so it transcribes

Where you drop a file decides whether it gets transcribed, and most complaints about a missing transcript come down to placement rather than the model. Descript only transcribes media that sits in the script, so a clip parked in the wrong lane stays silent text-wise even though it still spent media minutes on upload. Four rules cover almost every miss:

1. Drag the file into the script. That uploads and transcribes in one step; the Upload button in the Project panel only uploads.
2. Keep speech on the script track or in a sequence. A file added only as a layer is never transcribed.
3. For multitrack audio, insert the sequence into the script. A new sequence waits untranscribed in the Sequences folder until you do.
4. Split anything longer than 15 hours. Such files fail to transcribe.

Other imports transcribe on their own. Since March 2026, files added to the Drive Media Library are transcribed automatically and keep their transcript. A Zoom import adds the files to the script, builds a sequence, detects speakers and transcribes the audio, and it counts only the longest file against your minutes. Rooms recordings arrive as a transcribed sequence with multicam layouts applied. On import, Descript builds a lower-resolution proxy for speed, but the final export always uses the original. One catch: replacing a transcribed file does not regenerate its transcript, so run Transcribe again from the file's actions menu.

Choosing the language and the model

Descript lists 25+ transcription languages.^[3] Its help page calls the whole list Latin-alphabet, but that holds only for the fully supported ones: Greek and Hindi, which use other scripts, are in beta along with Welsh. Chinese, Japanese and Russian are not supported. Welsh must be picked by hand, because auto-detection skips it. Each file takes one language only, and files with several spoken languages are not supported, so split a bilingual interview before you transcribe it. If the language was wrong, re-transcribe the file from the Files section of the Project sidebar. Music and lyrics are skipped.

Since May 2026 the default model is ElevenLabs Scribe v2, and the Automatic setting routes every supported language to it. Scribe v2 spells in American English. For British spelling, pick English (UK) by hand in the language dropdown, because auto-detection will not choose it, and switch the model to Rev v2. Rev AI v3 joined the picker for everyone in August 2026. The settings live under App Settings, then AI models, where you can also set a default language or have Descript ask before every transcription. Spoken numbers and dates get formatted only in English, French and Spanish.

Raising accuracy before you correct anything

The transcription glossary is the cheapest accuracy fix. Add proper names, product names and jargon, and future transcripts recognize them more reliably. It holds up to 30 terms per Drive, is shared by every user and project on that Drive, and is available on all plans. It works in English only and does not touch translated transcripts. By default, Descript adds a word on its own after you correct it three times. On Business and Enterprise, the team manages one shared glossary inside Brand Studio. Cloudinary, in a customer story Descript published, credits the glossary for teaching the AI its technical terms.

Speaker detection runs on files up to 5 hours long. Automatic mode labels voices without asking how many there are; manual mode asks for the count first, set by the Always ask before detecting speakers option. The pricing page says it handles 8+ speakers, and Speaker Detective plays a clip of each voice so you can name it. Since June 2026, detection runs in the background on Drive media, so labels are ready right after import. Labels then follow the words as you cut and paste between compositions.

Mic bleed is the usual multitrack problem. One guest's voice lands on two tracks and the text doubles. You can drop a track from the combined script in the Sequence Editor, or run Fix duplicate transcripts, which removes duplicated text only and leaves the audio alone. Already have a professional transcript? Import transcript replaces Descript's version completely, but like the glossary it accepts English only.

Correcting and cleaning the transcript

Correct changes the words on screen without touching the audio or video, and Correct All fixes every instance across the project. Use it for misheard names. Editing is different, because deleting a word in the script cuts that moment from the media. Two softer options control what reaches the export. Ignore strikes the text through but keeps it, so it can appear in an exported transcript but not in subtitles or web-link media. Remove from transcript keeps the media but hides the text from the script, captions and exported transcripts.

Filler words are where free transcription ends. Since August 2026, detection works in English, Spanish, German, French, Portuguese and Italian, with the language detected automatically, and it spends AI credits, capped at 30 credits per file or application. On Free, the 100 credits are one-time, and three files at that cap use 90 of them. Each filler can be deleted, deleted with a gap, ignored, or removed from the transcript with the audio left as is. Silence and ambient noise can be labeled as wordless media, a beta feature that is off by default and switched on per project. A June 2026 fix stopped Scribe v2 from inventing words on near-silent files. Still, check the quiet stretches. Our Descript AI guide explains how credits drain across the other tools.

Exporting transcripts, subtitles and translations

A transcript exports as .html, .md, .docx, .txt or .rtf, with optional timecodes, speaker labels, markers and ignored text.^[4] Text exports leave out wordless media and scene boundaries. Subtitles are a separate .srt or .vtt file, with settings for speaker labels, characters per line and lines per card, and since June 2026 they carry word-level timestamps. Captions behave differently. They are styled from the script and stay on the video wherever it is shared or uploaded.

For YouTube, the corrected file is the whole point. YouTube adds its own English (Automatic) subtitle after upload, which you delete by hand before adding the SRT or VTT from Descript in YouTube Studio. Watch for doubling too. A local video export embeds subtitle cards, so a composition that also shows captions can display the text twice on some players. Publishing to a Descript web link avoids that, since files downloaded from a share page carry no embedded subtitles.

For automation, the API exports text, Markdown, HTML, RTF, DOCX or SRT, while the MCP server offers the same list minus DOCX. The Zapier action returns only text, Markdown, HTML or RTF. Translated captions cover 61 languages and export as txt, srt or vtt, singly or in batch; the help center and the pricing page disagree on which plans include translation, so check the plan before buying for it. Edits to the original script do not sync to translations, and re-translating overwrites manual fixes. Finish corrections first.

The bottom line

Descript gives you unmetered transcription in 25+ languages (Greek, Hindi and Welsh still in beta), a 30-term glossary, a choice of Scribe v2 or Rev models, and exports in every common text and subtitle format. The real limits sit around it: media minutes per plan, English-only import and glossary, one language per file, and AI credits for filler-word removal. Pick the plan by how many hours of new media you add each month, seed the glossary before your first long file, and correct before you translate or publish.