Descript can transcribe interviews and meetings you upload or record in it, label each speaker and export the text as Word, text or subtitle files. Descript says accuracy reaches up to 95% on clear audio, its production AI uses no customer data, and data sharing can be switched off; share links, retention and consent are yours to manage.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowThat 95% is a ceiling for clean recordings, not a promise for yours. On a real interview the errors gather where they do most damage: names, technical terms, numbers and people talking over each other. So the useful answer has two parts: run a ten-minute accuracy test on one of your own recordings before you rely on Descript, and change three privacy settings before the first upload.
How accurate is Descript's transcription on real interviews?
Descript's transcription reaches up to 95% accuracy on clear audio by Descript's own figure, and falls with background noise, accents, distant microphones, crosstalk and unfamiliar terms. Your own recordings decide where you land.
Descript's automatic transcription article says it uses third-party AI models it tests and monitors, and lists what moves accuracy: recording quality, accents, background noise, microphone placement, overlapping speakers and the transcription model. Several settings help before you transcribe anything:
| Setting or habit | Where | What it fixes |
|---|---|---|
| Transcription model | Settings, AI models, Transcription (under Core) | Some models handle punctuation, proper nouns or non-English speech better |
| Default transcription language | Settings, General, Transcription settings | Wrong-language transcripts from auto-detection |
| Always ask before transcribing | Same section | Lets you pick the language, and stops files transcribing straight after import |
| Transcription glossary | Project menu, Tools, Transcription Glossary | Names, brand terms and jargon (English only, up to 30 terms per drive) |
| Separate audio track per person | Your recording tool (for example Zoom's recording settings) | Crosstalk and speaker mix-ups |
Know the hard limits too. Descript transcribes one language per file; its help centre lists 26 languages (its pricing page says 25), all in the Latin alphabet, and a few, including one English variant and Welsh, must be chosen manually rather than auto-detected. It transcribes speech only, not music or lyrics. Files over 15 hours fail to transcribe automatically. Numbers and dates spoken aloud are formatted as digits ("nineteen forty-five" becomes "1945") in English and only two other languages so far.
Recording habits that matter more than any setting
Most transcription errors are decided before the file reaches Descript. Four habits make the biggest difference for interviews and meetings. Give each person their own microphone, even a cheap one, rather than one laptop microphone in the middle of the table. Ask remote participants to wear headphones: Descript's own guidance says mic bleed usually happens when a speaker without headphones is picked up by their own microphone and someone else's, which duplicates text and confuses speaker labels. Start every recording with each person saying their name and role, which gives speaker identification a clean sample and the transcript a record of who was present. And agree a simple rule for crosstalk in meetings, such as the chair naming the next speaker, because overlapping speech is the one thing no setting repairs afterwards.
On the glossary, Descript's own pages differ: the help article says it is on all plans, with shared team glossaries through Brand Studio on Business and Enterprise, while the pricing table lists the glossary under Business. Look for Transcription Glossary in your project's Tools menu; if it is there, your plan has it.
A veterinary practice's ten-minute accuracy test
Here is the test with illustrative numbers. A veterinary practice wants transcripts of two kinds of recording: recruitment interviews for veterinary nurses, and its fortnightly clinical meeting, where drug names and doses come up constantly. The practice manager picks a ten-minute stretch of a clinical meeting that includes the hardest parts: four speakers, a dose discussion and some crosstalk.
She makes a reference transcript by listening and correcting carefully (40 minutes of work for ten minutes of audio), then compares Descript's version word by word, counting three kinds of error: substitutions (wrong word), deletions (missed word) and insertions (extra word). The standard measure is the word error rate: errors divided by the number of words in the reference.
| Measure | First pass | After glossary and model change |
|---|---|---|
| Words in reference | 1,480 | 1,480 |
| Substitutions | 38 | 21 |
| Deletions | 12 | 10 |
| Insertions | 6 | 5 |
| Word error rate | 56 / 1,480 = 3.8% | 36 / 1,480 = 2.4% |
| Word accuracy | 96.2% | 97.6% |
| Drug names right | 5 of 14 | 12 of 14 |
| Doses and numbers right | 17 of 19 | 18 of 19 |
| Speaker labels right (paragraphs) | 41 of 47 | 45 of 47 |
The headline accuracy looked excellent on the first pass, yet 9 of 14 drug names were wrong, which is the number that matters to a clinical record. Adding the practice's 22 most-used drug and condition names to the glossary, choosing a different transcription model in Settings and using Re-transcribe on the file fixed most of them. The one dose still wrong, "0.2 ml" heard as "2 ml", is the kind of error no tool should be trusted with alone.
The decision that followed: Descript transcripts are fine as working notes and for finding moments in the recording, but any dose, drug or decision that goes into a clinical record is checked against the audio by the vet who said it. For recruitment interviews, the transcript is accurate enough to compare candidates' answers, with names and figures checked before anything is quoted back to a candidate.
If calculating the error rate and tracking test results in a spreadsheet is the slow part, try The Complete Excel Course with AI Integration on Coding Liquids.
How do you run the accuracy test on your own recordings?
An accuracy test for Descript takes about an hour: choose ten difficult minutes of a real recording, write a careful reference transcript, count errors against it, and check names, numbers and terms separately.
- Pick the hardest ten minutes, not the easiest: several speakers, the jargon you use, a phone participant if you have them.
- Write the reference. Listen at slow speed and type exactly what was said, or correct a transcript slowly until it is exact.
- Count errors into substitutions, deletions and insertions, and divide the total by the reference word count.
- Count critical items separately: names, numbers, product or drug terms, and any word that would change a decision if wrong.
- Fix what you can (glossary, model, language setting, separate tracks) and re-transcribe the same file to measure the change.
- Set your rule: what the transcript can be used for unchecked, and what always gets checked against the audio.
A filled-in test sheet for one recruitment interview, illustrative:
Recording: nurse interview, candidate B, 10:00-20:00, two speakers, office mic
Reference words: 1,260
Errors: S 19, D 7, I 3 = 29 Word error rate 2.3% Accuracy 97.7%
Critical items: candidate's name 3/3, previous employer 1/2,
registration number 0/1, salary figure 1/1
Speaker labels: 30/30 paragraphs right
Rule: transcript OK for comparing answers; registration number and
employer always checked against the application form.
Ten minutes is enough to see the pattern, because errors repeat: a name misheard once tends to be misheard every time. For a deeper version of this test on poor audio, see how accurate AI transcription is on poor audio, and for testing any tool on your own calls, testing transcription accuracy before you commit.
What does transcribing an hour of recordings cost in Descript?
Transcribing in Descript costs media minutes, not AI credits: used in full, a month's allowance works out at $1.60 an hour on Hobbyist, $0.80 on Creator and $1.25 on Business (annual billing).
Descript charges media minutes when you upload or record, and its transcription article says transcription itself is not separately metered. So the sum is simply hours of recording against the plan's monthly allowance, per editor, with nothing carried over at renewal.
| Plan (annual billing) | Media hours a month | Cost per hour if fully used | Extra hours |
|---|---|---|---|
| Free ($0) | 1 | $0 | None |
| Hobbyist ($16) | 10 | $1.60 | No top-ups; upgrade instead |
| Creator ($24) | 30 | $0.80 | Top-ups: 5 hours $25, 20 hours $80, 50 hours $150 |
| Business ($50) | 40 | $1.25 | Same top-ups |
The veterinary practice's month: two clinical meetings of an hour each and six recruitment interviews of about 35 minutes, so 120 + 210 = 330 minutes, five and a half hours. Hobbyist covers that with room to spare; the reason to pay more would be editing features or credits, not transcription. A busier month of 12 interviews would add another three and a half hours and still fit.
Two details keep the meter honest. A recorded meeting is charged once by its longest file, not once per participant, so a one-hour call with separate tracks for five people uses 60 minutes. And cancelled uploads are not charged, so stopping a wrong file mid-upload costs nothing.
How does Descript label speakers in an interview or meeting?
Descript detects distinct voices automatically during transcription (for files up to five hours), then asks you to name each speaker once from an audio sample, and applies those labels throughout the transcript.
- Import the file. Speaker detection runs in the background as part of transcription.
- When the notification appears, select Identify speakers.
- Play each voice sample and type the person's name, or choose an existing name.
- Select Close, then add the file to the script. Labels appear on each paragraph.
If detection is switched to manual in Settings, Descript asks how many speakers there are before it starts; files shorter than 10 seconds or longer than five hours cannot be detected. To fix a label later, place the cursor and type @ to choose a speaker, or use the menu beside a speaker's name to rename them or replace them throughout the project.
Meetings with everyone on one microphone are where labels go wrong most often. Two fixes help. Record a separate audio track per person where your tool allows it: Zoom has a setting for this, and Descript then builds a multitrack sequence with each person on their own track. And if two microphones picked up the same voice (mic bleed, usually from someone not wearing headphones), Descript can spot the duplicated text and offer a Fix button; it removes duplicated text from the script but not the audio, and it can be reverted from the file's Advanced menu if it removed the wrong thing.
How do you compare interview answers across candidates in Descript?
Compare candidates in Descript by highlighting each answer in a colour assigned to its question, copying the highlights of one colour into a new composition, and reading or exporting the answers side by side.
- Give each interview question a colour: say yellow for "Tell us about a difficult client", green for the clinical scenario, blue for availability.
- In each candidate's transcript, select the answer and choose the colour from the selection toolbar.
- Open the action bar with Command + K (Mac) or Ctrl + K (Windows), type "Copy highlights" and choose to copy one colour.
- Create a new composition called "Q2 clinical scenario, all candidates" and paste. Repeat for each colour.
- Export each composition's transcript with speaker labels, or play the answers back to back for the second interviewer.
Comparing the same question across three candidates in one place is fairer than flicking between three sets of handwritten notes, and it shows when one candidate was asked a longer or easier version of a question.
A summary can help, used carefully. An illustrative instruction for the Publish section's summary action, run on one candidate's interview: "Summarise this candidate's answers in five bullet points. Quote, don't judge. Note anything they said they could not do." An illustrative output:
- Three years as a veterinary nurse in a mixed practice.
- Described handling an anxious owner by explaining each step aloud.
- Clinical scenario: listed fluid therapy checks in the correct order.
- Strong team player who would fit well with the practice.
- Cannot work Saturdays before March.
What to fix: bullet four is a judgement the candidate never made and the panel never agreed, so delete it. Summaries of interviews should report what was said, with the judging left to the people accountable for the decision. Everything else checks against the transcript.
How do you correct and export a Descript transcript?
Correct a Descript transcript with Correct mode, which changes the text without touching the audio, then export it from Export, Transcript tab, as a .docx, .txt, .rtf, .md or .html with speaker labels and timecodes.
Correcting
- Double-click a wrong word, select Correct in the toolbar (or press C), type the fix, then choose Correct for this instance or Correct All for every instance in the project.
- For many fixes, press Option + C (Mac) or Alt + C (Windows) to enter Correct mode and type into the transcript like a document.
- Hold Z and click a word to toggle its capital; hold X and click to cycle its punctuation.
- Correct the same word three times and Descript adds it to the glossary automatically, a setting you can switch off.
Be careful not to confuse correcting with editing. Outside Correct mode, editing the script edits the recording: deleting a word there deletes its audio. Use Correct for spelling fixes, because it changes only the text.
Exporting
Select Export in the top-right corner, then the Transcript tab, choose a format and the options to include: composition name, markers, ignored text, speaker labels, or speaker names on every paragraph. Timecode settings add a time every set number of seconds, at paragraph breaks, or with speaker labels, which makes a meeting transcript easy to check against the recording. Copy to clipboard works for every format except .docx. For captions on a video platform, export subtitles instead, as .srt or .vtt, with an option to show speakers.
A transcript is not yet a record
A raw meeting transcript is long, repetitive and full of half-sentences, which is accurate but hard to use. Before and after for one item at the practice's clinical meeting (illustrative). Transcript, 140 words: "So the, um, the ear cytology, we've had, what, three this week where it came back, it came back mixed, and I think, I mean, I think we should probably, yeah, we should repeat at ten days rather than fourteen, if everyone's, does anyone, no? Okay, ten days then." Minute, 24 words: "Ear cytology: three mixed results this week. Agreed: repeat cytology at 10 days, not 14. Action: head nurse to update the protocol sheet." Keep the transcript and the recording as the evidence, and write the minute as the record. The timecodes in the export let anyone check the minute against the moment it came from.
One trap for meeting records: ignored (struck-through) text can be included in an exported transcript, while deleted text is not. If a remark was struck through because it should not be in the record, delete it rather than ignore it, or leave Include ignored text unticked.
How private are interview and meeting transcripts in Descript?
Descript says its production AI uses no customer data, research data sharing can be switched off, files are encrypted, and deleted projects are removed within 30 days; share-link access is yours to set.
The facts from Descript's security page and help centre, read for this tutorial:
- Training. Current production AI models use no Descript user data; internal research uses only data from users who opted into data sharing. Sub-processors such as OpenAI are under agreements not to train on the data.
- Transcription providers. Automatic transcription uses Rev, which Descript says deletes the data from its servers shortly after completing the job, and Whisper, an open-source model hosted in Descript's own infrastructure.
- Storage and encryption. Audio, video and transcripts are stored on Amazon S3 or Google Cloud, encrypted with AES-256 at rest and TLS 1.2 in transit. Descript states it is SOC 2 Type II compliant.
- Deletion. A deleted project sits in the drive's Trash for 30 days and is then deleted automatically; the Trash cannot be emptied early. Deleting your account removes all projects, drives and transcripts.
- Enterprise. Data sharing is off by default and cannot be switched on; the plan adds custom retention, single sign-on, SCIM and audit logs.
The three settings to change before the first interview goes in:
- Data sharing off. Settings (Cmd/Ctrl + comma), Profile tab, Share data with Descript: select Allowed to switch it off.
- Share page default. Drive owners can set the default privacy for every new web link. Public pages can appear in search engines, so set the default to Project access required for internal recordings.
- Transcribe on purpose. Turn on Always ask before transcribing if you sometimes upload recordings you have not yet decided to keep.
If you want help fitting transcription into how your practice or business actually runs, I offer one-to-one AI consultations for small businesses.
What should interviewees and meeting participants be told?
Interviewees and meeting participants should be told, before recording, that an AI tool records and transcribes the session, what the transcript is for, who sees it and when it is deleted. Then ask for their agreement.
Rules on recording consent and personal data vary by place, so take advice for anything sensitive, such as health information or recruitment records. The practical minimum is a short notice in the invitation and a spoken reminder at the start. A filled-in version for the veterinary practice's recruitment interviews, illustrative:
Before your interview
We'd like to record this interview and create a written transcript using
an AI transcription tool (Descript). The transcript helps the two
interviewers compare notes fairly. Only the interview panel sees it.
We delete the recording and transcript [8 weeks] after the role is filled,
unless you are offered the job, in which case they join your staff file.
If you'd rather not be recorded, tell us and we'll take written notes.
[your name], Practice Manager
Keep what you promise. If the notice says eight weeks, put a reminder in the calendar to delete the project, remembering that it then sits in Descript's Trash for 30 more days before it is gone. A transcript kept longer than promised is a bigger risk than one never made. For a fuller check of uploads, voice cloning consent and data, see whether Descript is safe for business uploads.
When is Descript the wrong tool for transcribing meetings?
Descript is the wrong tool when you only want automatic meeting notes and action items, when a meeting mixes languages, or when you need a certified verbatim transcript; it suits editing, clipping and reusing recordings.
| You need | Descript fit | Better route |
|---|---|---|
| Searchable transcript of an interview you will quote or clip | Strong | Descript |
| A short video recap of a meeting | Strong | Descript |
| Automatic notes and actions from every call, without editing | Weak: you upload or record, then work in the editor | A meeting note-taker |
| A meeting held in two languages | No: one language per file | A tool with multilingual transcription, or two passes on split files |
| Chinese, Japanese or Russian speech | No: Latin-alphabet languages only | A tool that supports those languages |
| A verbatim record for legal or clinical files | Draft only | Human transcription or human checking |
For the note-taker route, free AI transcription for client calls covers the limits and privacy of the free options. Where Descript fits, the routine is short: record with separate tracks, set the language, run the glossary, name the speakers, check the critical items, and export with timecodes so anyone can jump from a sentence to the moment it was said.
Questions about transcribing with Descript
Does transcription in Descript use up media minutes?
Uploading or recording does; transcription itself is not metered separately. Descript says media minutes are spent whenever you upload or record audio or video, while its transcription article says transcription is not tracked or metered on current plans. A one-hour interview therefore uses 60 media minutes when you upload it, and transcribing it costs nothing extra.
Can Descript transcribe a meeting held in two languages?
Not in one pass. Descript supports one transcription language per file, and files with several spoken languages are not supported. Its help centre lists 26 transcription languages, all written in the Latin alphabet, so languages such as Chinese, Japanese and Russian cannot be transcribed yet. A few, including one English variant and Welsh, must be chosen manually rather than auto-detected.
How long can a recording be for Descript to transcribe it?
Up to 15 hours. Descript says files longer than that fail to transcribe automatically, and automatic speaker detection works on files up to five hours long. The maximum upload size depends on the plan: 1 GB on Free, 10 GB on Hobbyist, 20 GB on Creator and 50 GB on Business. Split very long recordings before uploading.
Can I paste my own transcript into Descript instead?
Yes, for English. Descript lets you bring your own transcript into a composition, but its language support page lists importing and replacing transcripts as English-only features. That suits interviews already transcribed by a person, where you want Descript only for clipping and editing the recording.
What does Descript's Enterprise plan add for privacy?
Enterprise adds controls a regulated business may need: Descript lists opting out of training, custom retention, SOC 2 Type II documentation, single sign-on, SCIM user provisioning and audit logs. Data sharing is switched off by default on Enterprise drives and cannot be toggled on. Pricing is custom, through Descript's sales team.
Further reads
- AI vs Human Transcription: Which Is Worth Paying For? — When a human transcriber is still worth paying for.
- Descript for Business: Meeting Recaps, Client Updates and Training Clips — Turn the transcribed meeting into a recap or training clips.
- AI vs Human Medical Transcription: Which Should a Clinic Choose? — Clinical terms and transcription: how clinics choose.
- Descript vs Riverside: Which Should a Podcast Producer Use? — Descript against Riverside for recorded conversations.
- Which AI Tools Can a Small TV Production Company Use? — Choose a transcript and logging tool, build a searchable clip log, and keep human sign-off on rights.
- AI Dictation for Lawyers: Can It Replace the Dictaphone? — Compare AI dictation for a law firm, test it on your own legal vocabulary, settle confidentiality and set up a correction workflow.
- Is Descript Worth It? Which Plan a Weekly Show Needs — Four episodes a month needs far fewer media hours than most people fear. Which Descript plan a weekly show needs, with the sums.
Browse all AI-for-business tutorials
Sources: Descript help centre articles on automatic transcription, supported transcription languages, the transcription glossary, correcting transcripts, speaker detection, speaker labels, fixing mic bleed, transcript and subtitle export, app settings, account data and privacy, deleting projects and deleting your account; Descript pricing page and security page. All read 22 November 2026.
