How to quote an interview accurately from a transcript

A lot of misquotes start the same way: someone trusted the transcript. The words looked right, the sentence read naturally, and nobody went back to the tape. Then the person quoted writes in to say they never said that — and they're right.

Automatic transcription has made this easier to do and harder to notice. A machine transcript doesn't look like a draft. It has punctuation, paragraphs and speaker labels, and it reads like a record. It isn't one. This is a routine for getting from that draft to quotes you can defend.

Why a clean transcript is the dangerous kind

When a recogniser can't make out a stretch of audio, it usually doesn't say so. Sometimes it writes words that are plausible but wrong: a garbled phrase comes back as a grammatical one, a mumbled "I didn't" as "I did". Sometimes it leaves words out, and the sentence closes over the gap. Either way the errors that matter don't look like errors. The sentence around them is fine, so your eye slides straight over it.

So the question isn't whether the transcript is mostly right — it probably is. It's whether the particular sentence you're about to put in quotation marks is right, and nothing on the page can tell you that. Only the audio can.

Where the meaning-changing errors hide

Some mistakes are cosmetic. These aren't:

• Negatives. "Can" and "can't", "did" and "didn't", "was" and "wasn't" often differ by a sound that's barely there in fast speech. One flips the meaning of the whole quote.
• Numbers. Fifteen and fifty. Two hundred and two thousand. Percentages and dates, which people often say quickly and without emphasis.
• Names and jargon. An unusual surname, a company, a drug, a place — rare words are where a recogniser is prone to write down a more common word that sounds similar.
• Punctuation. The recogniser adds it. Where a comma or full stop goes can change who "they" refers to, or turn a question into a statement.
• Attribution. Speaker labels are a guess based on how voices sound. Short interjections and moments where people talk over each other are where a line lands on the wrong person — and a correct quote with the wrong name on it is still a misquote.
• The edges of a quote. Where a sentence starts and ends is also the machine's decision. A qualifier at the end of one sentence — "but that's only in the pilot" — can be split off into the next, and a quote lifted without it says something the speaker didn't.

The routine

1. Pick your quotes from the transcript, then stop trusting it. Use the transcript to find what you want to use. That's what it's good at. Then treat each chosen quote as unverified until you've heard it.

2. Listen to every quote at its own timestamp. Not a skim of the recording — the exact passage, from a few seconds before the quote starts to a few seconds after it ends. With a transcript where every word is timed, this takes seconds: click the line and you hear that moment. Listen for the negatives, the numbers, the names and where the sentence really ends.

3. Check who said it. For each quote, confirm the voice as well as the words. If speakers were labelled automatically, look hardest at short lines and at the moments on either side of a change of speaker.

4. Read the context. Play the lines around the quote. Was it a hypothetical? Were they quoting someone else? Did they correct themselves a sentence later? A quote can be word-perfect and still misleading if the context was doing half the work.

5. Clean up openly, or not at all. Removing "um", "you know" and false starts is common practice, but follow your publication's rules and be consistent. Mark cuts with an ellipsis, put your own words in square brackets, and never join two separate remarks into one quote as if they were said together.

6. Write down where each quote came from. A timestamp next to each quote in your notes costs nothing now and saves an argument later, when an editor, a lawyer or the person you quoted asks.

When the transcript itself has gaps

Sometimes the problem isn't a wrong word but missing audio: a phone call that dropped, a recorder that stopped when the screen locked. A transcript built from that recording can still read as if the conversation flowed straight on.

Before you quote across a gap, check whether there is one. In txscribe, a recording made in the browser that was interrupted says so in the transcript — where the audio stopped, and roughly how much is missing when that can be measured — and the note travels with the exports rather than staying on screen.

Quotes from another language

If the interview was in another language and you're quoting a translation, say so, and have the translation checked by someone who speaks the language before it's published. Machine translation of a machine transcript stacks two sets of guesses. And if the recording itself switched language, check that the passage you're quoting was transcribed in the language it was spoken in — automatic language detection can quietly translate instead.

Using AI answers to find quotes

Asking a transcript a question — "where does she talk about the budget?" — is a fast way to find material. It's also a new way to misquote, if the answer paraphrases and you take the paraphrase as the quote.

In txscribe, an answer lists the passages it cites underneath it, each linking to the moment it came from. A passage is shown in quotation marks there only when those exact words were found in the transcript; when a cited quote doesn't check out, you get at most the time, not words we can't stand behind. The answer's own sentences are a different matter: they are the model's summary, and quotation marks in them don't guarantee a verified quote — short phrases aren't checked, and neither is who said them. Quote from the passages, not the answer — and remember the transcript itself is still the machine's draft. Step 2 still applies.

The short version

The transcript tells you where to look. The audio tells you what was said. Every quote gets heard before it gets printed — with the right name on it, in its context, with its edges where the speaker put them.

If you'd like a transcript built for that kind of checking, you can try txscribe on one recording without an account. For attribution specifically, see how to transcribe an interview with speaker names.