clean-transcription

Clean Transcription

Two-stage cleanup:

1. Heuristic — regex pass over filler words, repeated stutters, all-caps shouting, and a per-project glossary of known mistranscriptions.
2. LLM polish (optional) — Claude rewrites cues for grammar/punctuation while preserving timing.

Output is a parallel file: <basename>.clean.srt (or .clean.txt). Original is never modified.

Procedure

1. Inputs

Field	Default
Source	required (.srt or .txt)
Mode	heuristic (default) / llm / both
Glossary	<project>/subtitles/glossary.json if present; or path supplied inline
Output	sibling <basename>.clean.<ext>

2. Heuristic pass

Filler words

Default list (case-insensitive, word-boundary):

um, uh, erm, ah, like, you know, sort of, kind of, basically, literally, I mean, right

Don't blanket-strip — like and right have legitimate uses. Strip only when surrounded by sentence-internal commas / mid-cue ("..., um, ..."). For SRT input, operate per-cue text without touching the index or timing lines.

# Per-cue text only — preserve "N\n00:00:01,000 --> 00:00:03,000\n..." structure
awk 'BEGIN{RS="\n\n"; ORS="\n\n"} {
  n=split($0, lines, "\n");
  for(i=3;i<=n;i++) {
    gsub(/\<(um|uh|erm)\>[ ,]*/, "", lines[i])
    gsub(/  +/, " ", lines[i])
  }
  for(i=1;i<=n;i++) printf "%s%s", lines[i], (i==n?"":"\n")
}' "$SRC" > "$OUT"

Stutters

Collapse adjacent identical short tokens: the the cat → the cat. Only collapse 1–4 character tokens (don't merge really really good).

Glossary

glossary.json is a list of {wrong, right} pairs. Apply as literal substitutions (not regex) unless explicitly marked.

{
  "substitutions": [
    { "wrong": "claud", "right": "Claude" },
    { "wrong": "ml flow", "right": "MLflow" },
    { "wrong": "kdenlive", "right": "Kdenlive", "case_insensitive": true }
  ]
}

Walk the array and apply with sed. For SRT, again confine to text lines.