Give every sound a purpose
Inventory music files, source audio, dialogue/reactions, sound effects and rights/context from the brief. Keep audio linked to source IDs and offsets. Do not fabricate words, recreate voices or infer a clean transcript without evidence. Captions reflect actual speech; leave uncertain words marked for review.
- Choose a music-led, selected-original-sound or dialogue-led structure. Do not run every camera microphone continuously under the song. Preserve short authentic laughs, doors, engines, arrivals or water when they contribute.
- Trim noisy starts/ends with short ramps to avoid clicks. Reduce noise conservatively, checking for watery artifacts. Use actual clean room tone when extending a pause; don't call unrelated sound authentic location audio.
- Automate music ducking around selected speech/reactions: start reduction just before the voice and recover smoothly afterward. Pick gain by intelligibility, not one fixed dB recipe. Check source channel layout, mono compatibility, clipping and strong level differences between clips.
- Use a J-cut when incoming scene sound begins before its image, or an L-cut when outgoing sound continues after the image changes. Keep story implications honest; don't imply events occurred together just because their sound overlaps.
- Put song changes at compatible phrases/chapter changes. Check rhythm, harmonic clash and volume during overlap. Use an intentional tail, fade or short ambience bridge when a beat blend sounds forced. Avoid abrupt endings and accidental double music.
- Attach SFX to visible impacts/reactions, not a generic swoosh at every cut. Respect humor=
none. Make sure a gag sound doesn't mask a word or overpower the music. - Render and measure the actual mixed file: integrated loudness, true peak, clipping, channels and discontinuities. Choose targets appropriate to the destination instead of treating one LUFS value as universal. Listen on available headphones/speakers and a small-speaker/mono preview when feasible.
Deliver an audio cue sheet with source, in/out, timeline positions, gain envelope, transitions and caption uncertainty. Record measured loudness/peaks separately from listening status and device. If listening was unavailable, state it and provide short review ranges. Numerical normalization does not prove speech is clear.