How to verify quotes in an AI transcript

Before you touch the transcript, copy the raw audio to a second location and treat that copy as the record of truth. Every later check — and any correction or complaint after publication — depends on being able to return to the unedited recording. If you recorded the interview with a phone or a recorder, transfer the file before you delete anything from the device. Your outlet's retention policy may set a minimum period for keeping interview recordings; when in doubt, keep it.

Verification is far faster when the transcript and the recording stay connected. Use a transcription tool where clicking a word plays the audio at that point and every line carries a timestamp — Notta is a strong example: its transcript editor keeps the audio synced to the text, and in our hands-on checks its transcript editing (correcting a misheard name and number, saving, and reopening the file) was the smoothest of the tools we tried, which matters when you are fixing errors against the audio. That editing step is part of verification, not separate from it: each error you correct while listening becomes a line an editor can later re-check at the same timestamp. Trint builds quote verification into its editor and exports verified quotes to its story builder; Otter.ai also synchronises playback with word highlighting. Whatever you use, the non-negotiable is the link between word and timestamp.

Select your candidate quotes from the transcript itself, not from the summary panel and not from an AI chat answer. Summaries and chat answers paraphrase, reorder and sometimes invent; the Mediahuis case in March 2026, where quotes in 15 of 53 posts did not appear in the source material and the journalist had written from AI-generated summaries, is the cautionary tale. Mark each candidate quote in the transcript and move to the next step with the audio open.

Start playback a few seconds before the quote begins so you hear the run-up, and listen at normal speed — slowed audio can mask how words actually run together. Read along with the transcript and correct the text word for word as you listen. This is where a synced transcript pays off: click the first word of the quote, listen, and fix. In our test, correcting a planted misheard name and number in the Notta transcript and saving the edit took seconds, and the correction survived reopening the file — exactly the behaviour you want when the transcript is your working document and the audio is your source.

Speaker labels are probabilistic. Trint's automatic speaker recognition, for example, only auto-assigns a speaker above 85% confidence and suggests a label below it — a useful reminder that labels are guesses, not facts. In a two-person interview you can usually tell the voices apart; in a multi-person interview, confirm each quoted passage by voice and by context, and check the label against your notes. If two people talk over each other, assume the label may be wrong until you have listened.

These are the highest-risk elements. Misheard names and unusual spellings, transposed numbers, dropped dates and — most dangerously — dropped negatives ("did not" becoming "did") change the meaning of a quote while sounding perfectly plausible. Research on speech-to-text hallucination (the Careless Whisper study) found invented phrases cluster around long pauses and silence, and older research on overlapping speech found recognition degrades considerably when speakers talk over each other. Listen to every name, number, date and negative twice, and give passages near pauses or crosstalk a second pass.

Lay your contemporaneous notes next to the transcript. Where they disagree, the audio decides. If a passage is genuinely unclear — heavy crosstalk, background noise, an inaudible word — do not fill the gap from the AI transcript or, worse, from a summary. Either paraphrase the point without quotation marks, mark the passage as unclear in your notes, or call the source back and confirm what they said. A callback is slower than trusting the machine, but it is the only option that produces a quotable, verifiable line.

For every quote you intend to publish, record in a quote log: the quote as verified, the timestamp in the original recording, the speaker, who checked it, and the date. This turns verification from a one-off act into an auditable trail. An editor — or a fact-checker responding to a complaint months later — can go straight to the timestamp and hear the passage for themselves. Keep the log with the recording under your outlet's retention policy.