Podcast transcript example: what a good one looks like

Here is the opening of a two-person interview episode, formatted the way most published podcast transcripts are. Read it first, then the reasoning underneath.

[00:00:04] HOST: Welcome back to The Long Version. I am Ade Fashola, and my guest today spent eleven years as a structural engineer before she started building furniture. Priya Raman, thanks for coming in.

[00:00:17] PRIYA RAMAN: Thanks for having me. Although I should say the eleven years is generous. The last two I was mostly making chairs in the evenings and pretending to be an engineer during the day.

[00:00:29] HOST: When did the pretending start to feel uncomfortable?

[00:00:33] PRIYA RAMAN: There was a specific Tuesday. I was in a meeting about load calculations for a car park, and I realised I had spent the entire hour thinking about a joint I could not get right on a bench. That was the moment I knew which one was the job and which one was the hobby.

[00:00:52] HOST: [laughs] And the car park is presumably still standing.

[00:00:55] PRIYA RAMAN: It is. I was good at the job. I just was not interested in it, and those are very different problems.

Timestamps sit at the start of each turn rather than every thirty seconds. This lets a reader find the audio for any line without cluttering the page, and it is the convention most podcast players expect if you ever convert the transcript to a chaptered format.

Speaker labels are in capitals and repeated on every turn, never abbreviated to initials after the first mention. Readers arrive at transcripts through search and land in the middle, so a label that only makes sense if you read from the top does not work.

The guest gets a full name, the host gets a role. That is a small thing that helps search: the guest's name is what people search for, and having it appear on every turn is honest rather than stuffed, because they did in fact say all of it.

Non-speech sounds are marked in square brackets and only when they carry meaning. The laugh is there because the line reads differently without it. Breathing, chair noise and the third um in a sentence are not.

Paragraphs break at turns, not at fixed lengths. A long answer gets internal breaks at natural pauses, which is a judgement call and the main reason human-reviewed transcripts still read better than raw machine output.

A verbatim transcript includes every um, every false start, every repeated word. A clean read removes them and lightly tidies grammar without changing meaning. The example above is a clean read, which is what almost all published podcast transcripts use.

Clean read exists because spoken language looks worse on the page than it sounds in the ear. Nobody notices a false start while listening. In print it reads as though the speaker was struggling, which is unflattering to a guest who did nothing wrong.

Use verbatim when the exact words are the point: legal, research, oral history, journalism where a quote may be challenged. Use clean read for anything published as a companion to an episode.

The line you should not cross is meaning. Removing um is tidying. Removing a hedge, a qualification or a change of mind is editing what someone said, and if a guest reads it later you want them to recognise themselves.

If your podcast sits on a site that has to meet accessibility standards, such as a university, a public body or a large employer, the transcript is not a marketing extra and the rules are stricter.

In practice that means every speaker identified, every meaningful non-speech sound described, and no content present in the audio that is missing from the text. Machine transcription alone generally does not clear this bar, because names, technical terms and crosstalk are exactly what it gets wrong.

The text also has to be real text on the page. A transcript inside an image, or locked in a PDF that was scanned rather than generated, is not accessible and is invisible to search at the same time.

If you are unsure which standard applies to you, ask whoever owns the website before you publish twenty episodes in a format that has to be redone.

Machine transcription is the sensible starting point for almost everyone now. Accuracy on clean single-speaker audio is high enough that editing the output is much faster than typing from scratch.

Where it still needs a human pass: proper nouns, company and product names, technical vocabulary, accents the model has seen less of, and any moment where two people talk at once. Those are also the highest-value words on the page, so the pass is worth the twenty minutes.

Recording quality does more for transcript accuracy than any choice of transcription tool. Two people on separate tracks, close to their microphones, in rooms with something soft in them, produces a transcript that needs almost no correction. One shared microphone in a hard room produces one that needs rewriting.

Hilite transcribes every episode as part of editing, which is how its text-based editor works, so the transcript is a by-product rather than a separate job. We make it, so weigh that as you like. Any transcription service will get you to the same place with an extra step.