clean-transcript

Clean transcript

Post-processing pass on an existing transcript to strip disfluencies.

Input

A transcript file, typically <stem>.raw.txt produced by transcribe-gemini-raw or transcribe-whisper-local.

What to remove

• Filler words: um, uh, er, ah, like (as filler), you know (as filler), I mean (as filler), sort of, kind of (as hedges only)
• Immediate repetitions: "I-I-I think", "the the"
• Obvious false starts: "I was going to— I wanted to say"
• Stutters and single-word restarts

What to preserve

• All content words and real intent
• Speaker's phrasing and vocabulary
• Paragraph breaks and any speaker labels
• Tone, register, and idiom
• Lists, enumerations, technical terms

What NOT to do

• Do not summarize
• Do not reorder sentences
• Do not add headers
• Do not paraphrase
• Do not correct grammar unless the raw transcript has a clear transcription error

Output convention

<source_stem>.cleaned.md in the same directory as the input transcript.

Implementation

This is an LLM task — read the input file, apply the rules above, write the cleaned version. No external tools needed.