How to Organize Interview Transcripts With AI Without Losing the Speaker

Keep the original recording, a raw machine transcript, and a separate working transcript. Do not correct the only transcript in place. The raw layer lets an editor inspect how names, pauses, interruptions, and uncertain words were originally captured; the working layer can add structure without pretending those changes came from the speaker.

Record interviewee, interviewer, date, location or medium, recording parts, transcript method, and any usage conditions supplied for the project. The Smithsonian Archives' oral-history guide recommends contextual introduction material about why the person was selected and the circumstances of the interview. That context also helps a future writer interpret the exchange.

Replace unstable labels such as Speaker 1 with confirmed names or neutral identifiers. Add timestamps at regular intervals and at every speaker change used for a quotation. Mark uncertain speaker assignments instead of guessing. If several recordings were joined, preserve the part number so a timestamp remains reproducible.

Correct obvious transcription errors in the working copy while keeping an edit log for material changes. Names, organizations, numbers, dates, and technical terms deserve a recording check. Use a consistent token such as [unclear 00:14:22] when audio cannot be resolved; a plausible completion is more dangerous than visible uncertainty.

Ask AI to propose topical boundaries using timestamp ranges and short neutral labels. Review those boundaries against the conversation. Keep the question that elicited an answer with the answer, and preserve interruptions or follow-ups when they change meaning. A quote about “it” may be unusable once detached from the object named thirty seconds earlier.

Use overlapping segments when one passage belongs to two themes, rather than duplicating and silently editing it. Keep a segment ID, start and end time, speakers, topic, one-sentence description, and relevant project question. The segment becomes a navigation aid, not a rewritten account of the interview.

Give the model a bounded set of segments and request themes with supporting segment IDs, exceptions, and alternative interpretations. A theme without a locator is a suggestion, not a finding. Ask which statements do not fit the proposed pattern; outliers often matter more than the most repeated phrase.

Separate descriptive themes—what topics recur—from interpretive themes—what those recurrences may mean. Keep the interpretation attributed to the writer unless the interviewee explicitly said it. If several interviews are compared, preserve participant identity and context rather than merging everyone into an anonymous composite voice.

A quote card should contain the exact transcript excerpt, speaker, recording and timestamp range, preceding question, short context, theme, edit status, and verification status. Replay the audio before accepting the quote. Check that cuts do not change tense, subject, certainty, or the relationship between question and answer.

If the publication style allows light cleanup, distinguish it from verbatim transcription and follow the interview project's editorial rules. Never let a generated paraphrase acquire quotation marks. Store paraphrases as writer notes with a link to the underlying segment.

Outline with themes and quote-card IDs. Use quotations where the speaker's language adds authority, character, or precision; summarize routine material in the writer's voice with accurate attribution. Keep contradictory or qualifying statements close enough that the reader is not given a false impression of certainty.

Before release, listen to the wider passage around every central quote and scan the full interview for later correction or withdrawal. Confirm that the planned use matches the permissions and editorial commitments governing the project. AI can reduce navigation time, but the writer remains responsible for representation.