dialogue_editing_adr v1.0.0

name: dialogue_editing_adr description: Edit production dialogue and record ADR that sits in the scene: clean and checkerboard dialogue, spot and cue lines for replacement, run the ADR session with beeps and streamers, match mic perspective, room tone, EQ and reverb, and fix lines at the syllable level. Triggers on clean up dialogue, dialogue edit, ADR, loop a line, replace a line, room tone fill, boom and lav mismatch, or dialogue that sounds pasted on.

Dialogue Editing and ADR

Purpose

Make spoken lines sound like they were recorded in the picture, on the day, by that person in that place. The work splits in two: first save every production line that can be saved, then replace only the lines that cannot be saved, and match the replacements so the cut between production and ADR is not audible.

Most bad dialogue work fails in one of three places. The editor cleans too hard and the voice turns thin and watery. The ADR is recorded too close and too dry, so it sits on top of the scene instead of in it. The edit cuts to silence between lines, so the room keeps vanishing and coming back.

Workflow

1. Build the dialogue session before editing a word. Duplicate the picture lock tracks and keep the original production audio muted underneath the working copy for the whole job. Split the sync sound into boom and lav tracks per speaker where both exist. Checkerboard scenes onto alternating track pairs so adjacent shots that sound different never share a track or a processing chain. Name tracks by speaker and mic, for example DX_MARA_boom_A, DX_MARA_lav_X. See references/dialogue-edit-workflow.md.

1. Spot the whole piece first, on headphones, at normal level. Play scene by scene and mark every problem with timecode, speaker, and a plain tag: noise, rustle, off-mic, overlap, mouth-noise, level, alt-take?, ADR?. Do not fix while spotting. The spotting pass produces the repair list, the alternate take hunt list, and the ADR cue list. That split matters: a line is only an ADR candidate after steps 3 and 4 fail on it.

1. Rescue from alternate takes before reaching for processing. Pull the same line from other takes, other angles, and wild lines. Steal a whole line first, then a phrase, then a word, then a syllable. Match waveform shape and mouth movement, not just the transcript. A real production word from a nearby take almost always beats a cleaned version of a damaged word and almost always beats ADR.

1. Clean in the lightest order that works. Work in this order and stop as soon as the line passes: cut and fill, clip gain, EQ, light repair, then noise reduction last.

• Cut out bumps, clicks, and rustle that fall between words and fill the hole with room tone from the same take, never with digital silence.
• Level single loud words and laughs with clip gain before any compressor touches the track.
• High pass for rumble. A starting range of 70 to 100 Hz suits many voices, set by ear per speaker.
• Use spectral repair in several light passes for clicks and short rustle on top of a word. Heavy single passes leave holes.
• Use noise reduction only on lines that need it, in small amounts, and A/B against the untreated line with music and effects playing. If the voice changes character, back off and accept some noise.
• Reduce mouth clicks and harsh breaths by hand where they distract. Leave natural breaths that belong to the performance.

1. Hold continuity across every cut. Lay a continuous room tone bed under the whole scene from tone recorded on the location, or build one from clean pauses in the same scene. Every edit point gets a short crossfade. Check three things at each cut: background level, background character, and voice perspective. If the background steps up or down, fix the bed. If the voice jumps closer or farther, that is a mic choice or EQ problem, see step 9. Never strip the gaps between lines down to nothing and then lay tone under it as an afterthought. That order is audible.

1. Write the ADR cue sheet from the spotting list. One cue per line, never per scene. Each cue carries cue number, timecode in and out, character, exact line as spoken on set (including words the script got wrong), reason for replacement, shot size, and whether the mouth is visible, partly visible, or off camera. Generate and validate the sheet with python3 scripts/cue_sheet.py spot.json --csv cues.csv --summary summary.md. Off camera and back-of-head cues are the cheap wins, do those first in the session. Visible close-up cues are the hard ones and need the most takes.

1. Set up the ADR session to match production, not to sound pretty. Ask the production mixer or read the sound reports for what was used on set. Record both mics when you can:

• Boom match: same type of shotgun or condenser as production, about 12 to 18 inches above the performer, angled at the mouth, mirroring the set position.
• Lav match: a lav on the performer about 8 inches below the chin, under the same kind of clothing layer if that is what production heard. Use a quiet, dead space. A treated booth is best. A closet full of clothes and duvets works and is a standard low budget answer. Keep the performer standing or sitting as they were in the scene, and get them moving for effort lines. A shouted line recorded sitting still in a booth will never match a body that is running on screen.

1. Run the session in loops with cues the performer chose. Work line by line in small chunks. Let the performer hear the original line, watch the picture, then record. Offer the cue style they work best with: three beeps before the in point, a single beep at the in point, a visual count-in, or a streamer wiping across the line. Fairlight in DaVinci Resolve and the ADR panels in Pro Tools and Nuendo all support beeps, streamers, and cue lists with take ratings. Record at least three usable takes of visible lines. Chase the emotion and the effort first, lip sync second. Direct for energy and intention, then fix timing in the edit. Note the circled takes on the cue sheet while you record, not from memory later.

1. Edit ADR for sync at the syllable level. Align by waveform and by mouth shape together:

• Line up the start of the line, then check each strong consonant, especially P, B, M closures, F and V teeth-on-lip shapes, and wide open vowels.
• Nudge single words or syllables rather than sliding the whole line when only part of it drifts. Cut on breaths and between words, hide cuts under narrow crossfades.
• Use time stretch tools such as VocALign or the equivalent in the DAW only after the manual nudge gets close. Stretch fixes timing, it does not fix a flat performance. If no take performs right, go back and record, do not stretch a dead take into sync.
• Keep the production line muted directly under the ADR so any word can be reverted in the mix in seconds.

1. Match perspective, tone, and space in that order.

• Perspective first: if production was a distant boom on a wide shot, the ADR must not sound like a close lav. Back the mic choice, level, and direct-to-reverb balance off before EQ.
• Tone second: EQ match by ear against the neighboring production lines. A common starting shape from ADR practice is cut below about 80 Hz for rumble, ease 200 to 400 Hz if the ADR is boxy, and add a small presence lift around 3 to 6 kHz only if the ADR sounds dull next to production. These are starting points, not settings. A/B on the actual lines.
• Space third: lay 10 to 30 seconds of the location room tone, looped cleanly, under the ADR at low level, then add short room reverb or a convolution reverb matched to the set, using early reflections to place the voice in the room size on screen. Automate reverb level per shot size: more room on wides, less on close-ups.
• If no clean tone exists, build it: pull clean pauses from the scene, or use an ambience match tool that learns the background from a clean section, or from the dialogue clip itself, and prints that background under the ADR.

1. Rebuild the life around the replaced line. ADR recorded in a booth has no breath effort, no clothing, no footsteps, no handling sound. Add back the production breaths where they still fit, plus Foley or library clothing movement, footsteps, and prop sounds that the original line carried. This layer does more to hide ADR than another EQ pass. If the production effects are usable, keep them and mute only the voice part of the original.

1. Check the scene the way a viewer hears it. Play the whole scene with music and effects at mix level, on speakers, without watching the waveforms. Mark any line you can identify as replaced without looking. Those lines go back to step 9 or 10. Then check on headphones for edit clicks and tone steps, and check the loud passages and whispers both survive. Soloing dialogue to approve it is how pasted-on ADR ships.

1. Deliver so the mixer can undo anything. Deliver the dialogue edit with production and ADR on separate tracks, ADR takes labeled by scene, cue, and take, each cue flagged replace or supplement, circled takes on top, alternates kept below, muted originals in place, and the room tone beds as their own tracks. Include the cue sheet with circled takes and a one page note listing lines that are still a compromise and why. Follow assets/delivery-checklist.md.

Output Contract

Deliver all of the following:

• Dialogue edit session: checkerboarded tracks by speaker and mic, muted production originals preserved, continuous room tone beds per scene, crossfades on every edit.
• Spotting list: every problem line with timecode, speaker, tag, and the route chosen: clean, alternate take, or ADR.
• ADR cue sheet: CSV from scripts/cue_sheet.py with cue number, timecodes, character, line as spoken, reason, shot size, mouth visibility, and circled take after the session.
• ADR takes: recorded boom and lav where possible, at least three usable takes for visible close-up lines, labeled SC<scene>_CUE<nnn>_T<take>.
• Match notes: per replaced line, the mic used, EQ moves, reverb type and level, and room tone source, so a mixer can adjust the match without guessing.
• Compromise list: lines that still read as replaced on a full scene playback, with the next fix you would try. A delivery with no compromise list on a hard scene is a delivery nobody checked.
• Delivery checklist: completed copy of assets/delivery-checklist.md.

Operating Rules

1. ADR is the last resort, not the first tool. Alternate take, then cleaning, then ADR. Every unnecessary replaced line is a line that now has to be matched, and matching costs more time than rescuing.
2. Never cut dialogue to digital silence. Silence between lines reads as a hole in the world. Fill with room tone from the same scene, every time.
3. Match performance before sync, and sync before processing. A well performed line slightly out of sync can be nudged. A perfectly synced flat line cannot be fixed with plugins.
4. One mic perspective per angle, held consistently. Do not run boom and lav together by default just because both exist. Pick the mic that carries the angle, park the other as a muted alternative, and intercut only after EQ matching. Mixing both without phase alignment smears consonants.
5. Clean less than you think. Noise that vanishes under music and effects at mix level does not need removing. Processing artifacts do not vanish. Judge cleaning in the full scene, not soloed.
6. Do not invent settings. EQ points, reverb times, and noise reduction amounts in this skill are documented starting points from practice. The right value is set by A/B against the neighboring production lines on that project, and the value used goes in the match notes.
7. Record more tone than you think you need. On any set or session you control, capture at least 30 seconds of clean room tone per setup and wild lines for problem lines on the day. Tone recorded later from a different room will not match.
8. If the mouth is clearly visible and no take syncs and performs, say so. Offer the honest options: another session, a cutaway or reaction shot over the worst part of the line, or keeping the production line with heavier cleaning. Do not ship a visibly out of sync close-up and hope nobody notices.