Transcript cleaner
What it does
Cleaning a transcript is the step between the recording and the analysis. Automatic speech recognition hands you guessed speaker labels, one person split across four of them, filler at full density, and sentences broken wherever the model heard a pause. Qualitative methods have names for the choices you make next: verbatim transcription keeps every stumble and repetition for close linguistic reading, while clean verbatim, also called intelligent verbatim, drops the disfluency and leaves the words intact. That choice decides what you're allowed to quote later, so it belongs to you.
This skill takes the export as it comes out of Otter, Fireflies, Zoom, Granola or a plain recorder, and returns a document laid out for analysis. It resolves speaker labels from the content of the conversation, so a moderator tagged as three different speakers becomes one moderator. Fragments get merged back into sentences, and timestamps stay at paragraph level so a quote can be traced to the recording. Anything the model couldn't hear is marked in place. Redaction is a switch you set before it runs.
What you get back
- A clean transcript with one label per person, in the order the conversation happened.
- Paragraph-level timestamps, so any quote can be found again in the recording.
- Filler and false starts removed at the level you choose, anywhere from full verbatim to clean verbatim.
- Inaudible and low-confidence passages marked in place with their timestamps.
- Optional redaction of names, employers and anything else you list.
- A header block carrying the date, the participant label, the moderator and the session length.
How to use it
Example Prompts:
"Clean these eight Otter exports and fix the speaker labels, P3 keeps getting split in two."
"Clean verbatim, keep the timestamps, redact company names."
"This Zoom transcript has no speaker labels at all. Work out who's who from the content."
What it won't do
It won't recover audio the recorder never caught. Where the model heard nothing you get a marked gap and a timestamp to go back to. It also won't quietly correct a misheard word into what it assumes you meant. Product names and unusual surnames come back flagged for you to confirm, because a confident wrong fix does more damage than an obvious one.
Questions
Which export formats does it handle?
The exports from the automatic transcription services researchers actually use, Otter, Fireflies and Zoom among them, plus text pasted straight in. Anything it can't parse comes back with an explanation of what it choked on, so you can convert the file and try again.
Will it change what people said?
No. Cleaning is limited to structure, speaker attribution and disfluency at the level you set. Word choice is preserved as spoken, and any passage it had to guess at carries a flag where it sits.
Does it work on focus groups and multi-person sessions?
Yes, though it gets harder when people talk over each other. It labels what it can from the content, and marks the stretches where attribution is uncertain instead of assigning them to whoever spoke last.
