Transcription Cleanup Rules

A transcript's job is to represent what was actually said in a form a reader can use — never to represent what the speaker "meant to say" or what reads better in isolation. Every cleanup decision below trades some literal accuracy for readability; make that trade deliberately and consistently, never word by word on a whim.

Verbatim vs. clean-read

• Decide up front which mode the transcript needs, and say so in a header note if it's ambiguous: verbatim (every "um," false start, and repetition preserved — used for legal, research, or accessibility-compliance transcripts) or clean-read (filler and false starts removed, grammar lightly smoothed — used for articles, show notes, and most captions).
• Never mix modes within one transcript — a paragraph that's aggressively cleaned next to one that's verbatim reads as inconsistent and makes the document untrustworthy as a record.
• Default to clean-read for anything meant to be published or read as prose; default to verbatim for anything that will be quoted as evidence, cited legally, or used to train/evaluate a model on real speech patterns.

Filler removal

• Remove filler words ("um," "uh," "like" used as a verbal tic, "you know" as a placeholder) only in clean-read mode, and only when removing them doesn't change the sentence's meaning or rhythm-bearing emphasis.
• Never remove a filler word that's doing actual work — a hesitation before a hard admission, or "like" used as an actual comparison ("it felt like a trap") stays.
• Collapse false starts and self-corrections to the corrected version only in clean-read mode ("I think we should— we definitely should ship this" becomes "We definitely should ship this"). In verbatim mode, preserve both the abandoned start and the correction.
• Don't remove filler so aggressively that a speaker's natural voice disappears — clean-read means readable, not robotic; a light touch beats an over-edited transcript that no longer sounds like a person.

Speaker labels

• Label every speaker consistently throughout — same name or identifier every time, never switching between "John" and "Speaker 1" for the same person mid-document.
• Use real names when known and appropriate for the context; use consistent generic labels (Speaker 1, Interviewer, Guest) when names aren't available or confidentiality requires it — pick one convention and hold it for the whole document.
• Start a new labeled line every time the speaker changes, even for a one-word interjection ("Right." / "Exactly.") — don't fold a second speaker's words into the first speaker's paragraph.

Punctuating spoken rhythm

• Use periods, commas, and dashes to represent how the sentence was actually paced, not to force spoken language into formal written grammar — a speaker's run-on thought can stay a long sentence with dashes rather than being chopped into unnaturally short formal sentences.
• Use an em dash for a genuine interruption or abrupt self-correction, and an ellipsis for a trailing-off or unfinished thought — don't use them interchangeably.
• Add paragraph breaks at natural topic or turn shifts, not at arbitrary line-length intervals — a transcript with no paragraph structure is as unreadable as one that's over-chopped.

Inaudible, crosstalk, and non-verbal conventions

• Mark unintelligible audio as [inaudible] or [inaudible 00:14:32] with a timestamp if precision matters — never guess at a word and present the guess as fact. If reasonably confident but not certain, mark it [word?].
• Mark simultaneous speech as [crosstalk] rather than trying to interleave two speakers' words in an order that implies a false sequence.
• Mark significant non-verbal audio relevant to meaning — [laughs], [pause], [phone rings] — only when it affects interpretation of the dialogue; don't annotate every breath or throat-clear.

Timestamps

• Insert timestamps at consistent intervals (e.g., every speaker change, every 30–60 seconds, or every paragraph break) matching the transcript's intended use — captioning needs frequent timestamps; an article-style transcript needs far fewer or none.
• Format timestamps consistently throughout ([00:14:32] or (14:32)) — never switch formats mid-document.

Never invent words

• If the audio is genuinely unclear, mark it as inaudible or uncertain rather than filling in a plausible-sounding word — an invented word presented as transcribed speech is a fabrication, not a cleanup, regardless of how minor it seems.
• When cleaning up grammar in clean-read mode, don't add words the speaker didn't say to "complete" a thought — restructure using only their actual words, or leave the sentence as a natural fragment if that's what was said.