Video Transcript Best Practices for Publishers
The decisions that separate a transcript people can rely on from a wall of machine output nobody trusts.
Should a transcript be verbatim or cleaned up?
Go verbatim when the exact words carry weight, and clean up when the ideas are the point. Those are two different documents serving two different readers, and trying to produce one that does both gives you something that fails at both.
Verbatim keeps every word, including repetitions, false starts and fillers. It's right for legal and compliance records, research interviews where speech patterns are data, journalism where a quote may be challenged, and accessibility transcripts, where a deaf or hard-of-hearing reader deserves what a listener gets.
Cleaned up removes fillers and stumbles, repairs sentence boundaries and leaves the meaning untouched. It's right for published interviews, show notes, documentation, page text and anything a person will read rather than audit.
Decide before you start editing, not halfway through. And say which one it is at the top: a line reading "lightly edited for clarity" tells readers how to treat what follows and costs you nothing.
How should speakers be labelled?
Label every speaker by name on their first turn and consistently after that, using the same form of the name throughout. Switching between "Dr. Lin", "Mei" and "ML" in one document makes it impossible to search or split programmatically.
Rules that hold up across formats:
- Put the label at the start of the line followed by a colon, with the speech on the same line or the next one. Pick one and keep it.
- Introduce an unfamiliar speaker once with a short role note, then drop to the plain name: "Mei Lin, product lead:" on the first turn, "Mei:" after.
- Use a stable placeholder for unidentified voices, like "Speaker 2", rather than guessing. A wrong name is worse than an honest unknown.
- For audience questions in a recorded talk, "Audience member:" beats inventing identities.
- Add a new label every time the speaker changes, including one-word interjections.
Automatic speaker separation is useful as a first pass and unreliable at turn boundaries, especially where people talk over each other. Assume you'll fix the handoffs by hand.
What should I do with filler words, false starts and crosstalk?
In a cleaned-up transcript, cut fillers and false starts silently; in a verbatim one, keep them and mark the interruptions. The mistake is doing it inconsistently, so some pages read as polished prose and others as raw output.
For cleaned-up work: remove "um", "uh", "you know" and "like" where it's filler rather than comparison. Drop abandoned sentence starts and keep the completed thought. Don't rewrite grammar, don't tighten phrasing into something the speaker didn't say, and never add words to improve an argument.
For verbatim work: keep the fillers and repetitions, and mark non-speech sounds that carry meaning in square brackets, such as [laughs] or [long pause]. Use them only where the sound changes how the words read.
Crosstalk is the hardest case either way. When two people speak at once, give each a labelled line in the order they started and mark the overlap with [crosstalk] rather than inventing a clean sequence of turns. If a passage is unrecoverable, say so instead of guessing.
How many timestamps does a transcript actually need?
Timestamp density should match how people will use the document: roughly every speaker change for an interview, every thirty to sixty seconds for a monologue, and not at all for a short clip under two minutes.
Timestamps earn their place when the reader needs to get back to the audio: verifying a quote, clipping a segment, navigating a two-hour recording. They become clutter when the transcript is read as prose, because a code every few lines breaks reading rhythm for no benefit.
- Use one format throughout. Pick [00:12:34] or 12:34 and never mix them.
- Include hours only when the recording runs past an hour.
- Put the timestamp at the start of a turn or paragraph, never mid-sentence.
- Keep a timestamped master even when you publish a clean version. Stripping stamps later is trivial; adding them back means redoing the work.
For long recordings, chapter markers at topic changes serve readers better than dense per-minute stamps, and they're far easier to scan.
How do I handle inaudible passages, uncertain names and numbers?
Mark uncertainty visibly rather than guessing, because a confident wrong name does more damage than a visible gap. Readers forgive a marked gap; they don't forgive discovering the transcript invented something.
Conventions most readers already recognise:
- [inaudible 00:14:22] for speech you can't make out, with the timestamp so someone else can try.
- [sp?] or [phonetic] after a name you heard but can't verify the spelling of.
- [unclear] for a passage where you caught the words but not the meaning, such as a reference to something off-camera.
Proper nouns are where machine transcription fails most reliably, and they're what readers notice. Before publishing, check every personal, company, product and place name against a source outside the audio. A speaker's own website settles most of them in seconds.
Numbers and acronyms need one house rule applied consistently. Write numbers under ten as words and use digits above, or use digits throughout for a data-heavy transcript, but don't alternate. Expand an acronym on first use with the short form in brackets, then use the short form.
Is a transcript the same thing as captions or subtitles?
No. A transcript is the full text as a standalone document, captions are timed text synchronised to video for viewers who can't hear the audio, and subtitles are timed text that translates dialogue for viewers who can hear but don't speak the language. They're built differently and they're not substitutes.
- Captions include non-speech audio that matters to comprehension, such as [door slams] or [ominous music]. Subtitles usually don't, because the viewer can hear it.
- Captions are broken into short timed cues sized for on-screen reading. A transcript runs in full paragraphs.
- A transcript can be read without the video at all, which is why it's the accessible artefact for screen-reader users and skimmers alike.
This is why a caption file is not a transcript. Convert an SRT to plain text and you get speech chopped into two-line fragments with no speaker labels and no paragraph structure. It's a starting point: you still have to rejoin lines, add attribution and restore punctuation. Full breakdown in captions versus transcripts versus subtitles, with the obligations in video accessibility and WCAG.
Should I export TXT, SRT or VTT?
TXT for anything a human reads, SRT for uploading captions to a video platform, VTT for captions on a web page you control. The choice follows the destination, not preference.
- TXT is plain text with no timing. Use it for show notes, articles, documentation, search indexing and quoting. Everything opens it.
- SRT is the most widely accepted timed-caption format. Use it when uploading a caption file to a video host or handing one to an editor.
- VTT is the web standard used by HTML5 video. Use it when serving video from your own site, and when you want chapters, which SRT doesn't carry.
Keep the timed file as your master and generate plain text from it, not the other way round. Converting SRT or VTT to TXT is cheap; going back means re-timing from scratch. Details in SRT vs VTT vs TXT.
What should I check before publishing, and whose content can I transcribe?
Run a fixed checklist before anything goes live, and keep the difference clear between transcribing for reference and republishing someone else's words.
- Spot-check the audio against the text at three or four random points, including the last few minutes, where machine accuracy tends to drift.
- Verify every proper noun against an external source.
- Confirm speaker labels are consistent and no turn is attributed to the wrong person.
- Confirm timestamp format is uniform and the stamps still line up after editing.
- Decide whether anything sensitive belongs in a public document: a phone number, an address, an unreleased detail, a third party named without their knowledge.
- Add one line stating whether the transcript is verbatim or edited, and the recording date.
On rights. Transcribing publicly available content for your own notes, research, search or accessibility is normal and widely practised. Publishing a full transcript of someone else's video as your own page republishes their work, and machine transcription doesn't change who authored the words. Quote the part you need, attribute speaker and source, add a timestamp, link back.
Private recordings are a harder line. Don't publish a transcript of a conversation that wasn't recorded for publication without the participants' agreement, and check your local recording-consent rules before relying on one. Tools like JotVox make the mechanical part trivial, which leaves the judgement as the only part that's actually yours.
Keep reading
Caption vs transcript vs subtitle, SRT vs VTT vs TXT, Video accessibility and WCAG, How accurate is AI transcription?.