SrigadaAkshayKumar/ai-video-editor

clean-cut

Goal: the viewer hears one fluent, correct take of everything the speaker meant to say, in order, with natural rhythm — and nothing else. Be ruthless with mistakes, careful with meaning.

Inputs

• work/transcript.md — lines L012 [mm:ss–mm:ss] w120–w134 | text, pauses, ⁽?⁾ low-confidence words, [events], long silences. Word ids are what you cut with.

• work/words.json — the same words with exact times and conf (0–1).

• Check the header's engine. whisper (English) is primed to transcribe verbatim and its word times are re-anchored to measured silences, but it can still smooth over a mumbled filler or a very fast stutter — treat an unexplained short gap inside a sentence, or a ⁽?⁾ word, as a possible hidden disfluency, and use a time cut if silences.json and the context confirm it. ElevenLabs (Telugu/Hindi) timings are tighter and it tags [events].

• work/script-diff.md — only when the user supplied a script (faceless): what was said vs what was written. Every "extra spoken run" is a cut candidate (re-read, false start, stumble, ad-lib); the suggested textFixes correct caption spelling from the script. An ad-lib that genuinely adds something may stay — the script is the intent, not a cage.

Read the entire transcript before writing a single cut. Retakes are only obvious in context.

What to remove

Work through these in order. Every removal gets a short reason.

1. Retakes — judged by MEANING, not by words. This is the part most often missed. A speaker re-explains the same point with different wording, a different example, a hedge first and the full claim later, or the same clause switched between Telugu/Hindi and English — with zero shared words. Ask "if I kept only one of these, would the viewer miss any new information?" If no, it is a retake: keep the ONE take spoken completely through to the end of the sentence (usually the last) and cut every other attempt, even non-adjacent ones, even ones that sound fluent in isolation. Be aggressive: do not keep an earlier restatement "for context" — the complete version reads fine alone. Keep a restatement only if it adds genuinely new information, or it is deliberate rhetorical emphasis in a tight couple of sentences (not a restart after a pause or hedge).

   Mandatory sweep before writing the EDL (fumble-and-respeak is the dominant pattern and the most common review complaint — the template's speakers attempted single claims 5-9 times):

   • Pass 1, chronological: for EVERY line ask "is this idea stated again anywhere later?" — restatements are routinely 2-6 lines apart with a silence, hedge or aside between.
   • Pass 2, by claim: list the distinct claims the video makes; any claim covered by more than one line is a retake cluster — keep exactly one.
   • Pass 3, inside each kept line: a single transcript line can hold a fumble — a phrase, a stall, the same phrase again ("akkarleni courses, akkarlenanta you know … akkarleni courses"). Passes 1-2 compare lines and cannot see it.

   Signals an attempt was abandoned (cut it even with no word overlap): it ends mid-sentence, mid-word, on a dangling connective or trails off; it is followed by a >1.5s silence and the same thought started over; it repeats the previous line's opening clause before going further (keep the run that gets furthest); the same sentence is tried 3+ times — keep only the last complete one.

2. False starts / abandoned sentences. Speaker starts, trails off or stops, restarts differently. Remove from the first word of the abandoned attempt to the start of the restart.

3. Stutters and doubled words. this is, this is → keep the last; the the → one. Cut the earlier copy so the kept word flows into what follows.

4. Fillers (only when used as filler, not as content):

   • English: um, uh, er, ah, hmm, like, you know, I mean, so (sentence-initial), basically, actually, right?, okay so
   • Hindi: मतलब/matlab, तो/toh (leading), हाँ/haan, अच्छा/accha, यानी/yaani, वो/woh (hesitation), ना/na, है ना/hai na, बस/bas, एक्चुअली
   • Telugu: అంటే/ante, ఆ/aa, ఏమో/emo, అదే/ade, మరి/mari (leading), ఇంక/inka, కదా/kada (tag), అలా/ala (hesitation), ఏంటంటే/entante, సో/so
   • Code-mixed speech mixes all of the above; ElevenLabs may write English words in Telugu/Devanagari script (e.g. సో, एक्चुअली) — treat them the same. Leave a filler in if removing it would make the cut audibly choppy and it is short — but default to removing.

5. Mispronunciations and slips. Signals: ⁽?⁾ / conf < 0.5, garbled ASR output, a word immediately said again correctly (Hyderbad— Hyderabad), a name said two ways. Remove the wrong attempt and keep the correct one. If a word is mispronounced and never said correctly, keep it and list it under "couldn't fix" in your report.

6. Inconsistencies / self-corrections. in 2019 — no sorry, 2020 → keep only the corrected statement (in 2020); remove the wrong fact and the correction phrase (no sorry, sorry, I mean, ఆ కాదు, नहीं नहीं). If two takes state different facts and neither is corrected, keep the last and flag it to the user.

7. Off-script material. "Is it recording?", "cut that", "let me do that again", "one more time", countdowns, talking to someone off camera, clapping/slates, coughs, sneezes, long breaths, lip smacks, phone noises, [events] that aren't content (keep genuine laughter that's part of the content). Use { "time": "a-b" } cuts for non-speech noise that sits inside kept speech.

8. Dead air — handled for you: lead-in/tail are trimmed and every pause is capped at keepPauseMax. Don't write cuts for ordinary pauses.

Voiceovers (faceless edits)

A narration track, read from a script, in a quiet room:

• Dead air between takes is the biggest win: keepPauseMax 0.3-0.35. Narration with air between every sentence sounds amateur.
• But keep a beat before a reveal: 0.35-0.55s before a line that lands a reveal, a number or a chapter turn — "holds": [{ "after": "<last word id before the beat>", "max": 0.5 }]. Say "held beat" in your notes; the next stage puts a graphic or the music gap there.
• Retakes: the narrator re-reads until clean — keep the last complete, best-read take.
• Reading artifacts: false starts, a swallowed word retaken, page turns, lip smacks, take slates, "okay"/"let me try that again" — cut them all; they are not in the script.
• Don't cut every breath (it sounds synthetic) — only breaths inside gaps you're removing anyway or a gasp before a retake.
• Keep the narration's rhetorical questions to the viewer ("ఇది మీకు తెలుసా?") and flag them — they become involvement beats.

What NOT to do

• Don't change meaning, merge two different points into one, or drop qualifiers ("not", "only", "almost", Telugu కాదు/లేదు, Hindi नहीं/मत) — re-read every join.
• Don't cut a word in half or leave a sentence without its verb (Telugu/Hindi are verb-final: the verb is at the end — make sure the kept take includes it).
• Don't reorder content (not supported; tell the user if a reorder would help).
• Don't remove deliberate dramatic pauses — if one matters, leave the words around it alone and raise keepPauseMax only if the whole video is meant to be slow.

edl.json

{
  "remove": [
    { "words": "w6-w12",  "reason": "false start; full retake at L003" },
    { "words": "w40-w58", "reason": "retake 1 of intro (kept take 2 at w59-w77)" },
    { "words": "w23",     "reason": "filler (um)" },
    { "words": "w90-w91,w95", "reason": "filler ante + stutter" },
    { "time": "83.20-84.05", "reason": "cough inside kept sentence" }
  ],
  "keepPauseMax": 0.35,
  "edgePad": 0.08,
  "textFixes": { "w26": "This", "w311": "HyperFrames" },
  "holds": [ { "after": "w88", "max": 0.5 } ]
}

• words ranges are inclusive, comma-separable; time ranges are raw-timeline seconds.
• keepPauseMax: 0.25–0.3 fast social cut · 0.35 default · 0.5 calm/educational.
• edgePad: room kept around each cut so consonants aren't clipped; 0.06–0.1.
• textFixes: caption text only (audio is untouched) — fix ASR misspellings, brand names, capitalisation after a cut, wrong script for English loanwords if the user wants Latin script.
• holds: keep up to max seconds of pause after a word (a dramatic beat), overriding keepPauseMax.

Process

1. Read the transcript; list the takes for each passage in your head (or in work/cut-notes.md for long videos) — "intro: L001–L004 (take 1, broken), L005–L007 (take 2, good)".
2. Write work/edl.json.
3. npm run cut -- <p> --dry-run → read the clean script it prints, line by line, as the viewer would hear it. Look for: leftover repeats, joins that break grammar, missing verbs, dangling "so"/"and"/"అంటే"/"तो", lost meaning. Fix the EDL and dry-run again until it reads cleanly.
4. npm run cut -- <p> → renders work/cleancut.mp4 + re-timed work/cleancut.words.json (shared by the 16:9 and 9:16 edits).
5. Read work/cut-report.md: every removal should be one you intended, with its reason, and the "clean script" should match your dry-run. The printed rendered duration must equal the expected duration (± 0.05s); if not, something went wrong in the encode — re-run.
   • Boundary audit: the cut measured every join on the real audio and moved any edge that caught the attack of the next (removed) take or the tail of a removed word — STT word times are routinely 50-80 ms off, so a word-timestamp check alone misses audible leaks. Each move is listed.
   • Fragile words: a very short or zero-width word sitting on a cut (in the template a zero-width "కాదు" on a boundary turned "not everyone gets them" into "everyone gets them"). Re-read each one's sentence; the better fix is usually to extend the kept segment to swallow the redundant line after it, removing the splice entirely, rather than nudging a boundary by milliseconds.
6. Report to the user in a few lines: raw → clean duration, % removed, count by reason (retakes / fillers / stutters / slips / off-script), and any "couldn't fix" items.