music-under-narration

Public-radio and feature-documentary practice for music against speech, made into numbers. OpenDubStage tool: opendubstage carve (opendubstage.mix.carve: carve, posts, phrase_safe_end, band_energy_ratio_db).

1. Spotting: music marks shifts

2. Entries, posts, exits (numbers)

Shorten by removing whole 4/8/16-bar phrases; crossfade one beat (≈ 200–600 ms at 60–120 bpm), equal-power.

3. Ducking: depth and frequency

The one controlled study: Torcoli et al., AES 146 (2019): listeners want music ≥ 10 LU under commentary and ambience ≥ 15 LU; non-experts prefer ~4 LU more; voice-over-voice ≥ 20 LU. The opposite fault: "if the music drops 15–20 dB on every line the audience hears the automation" (IRPR). WCAG 1.4.7 (AAA) is 20 dB and is a ceiling, not a target.

The pro move is a moderate broadband duck plus a spectral pocket where consonants live (TrackSpacer, Neutron Unmask, pure:unmask do this live; opendubstage carve does it offline).

opendubstage carve MUSIC.wav DIALOGUE.wav OUT.wav [--cut-db -6] [--band 1000 4000] [--gate-db -45] [--block-s 60]

Apply the envelope as a gain curve (10 ms frames, amultiply or volume=…:eval=frame), not a compressor; keep sidechaincompress at most as a mild safety net (ratio 2, release 1500). Deterministic, plottable, testable.

Depth is per song, not per film: a sparse piano pad can sit at −8; a dense vocal song under speech needs −14 to −17, or the instrumental. Sometimes the answer is an exit, not a duck.

4. The music budget

5. Nothing sung under speech

• No vocal section overlaps speech for > 1 s. Either swap the whole cue for its separated instrumental (BS-RoFormer via audio-separator, whole song at once), or run the vocal version in the clear and crossfade to the instrumental 0.5 s (one bar, equal-power) before the first word (the TAL way of using a song with words).
• Separate once per song used in under mode and cache (≈ 2–4× real-time on Apple silicon).
• Generated underscores must be checked for hidden vocals. Measured in practice: an LLM listener caught buried voices in a generated instrumental that Whisper and the generator's own lyric channel both missed. Procedure in ai-review-of-audio (references/prompts.md, "hidden vocals").
• Lyric as button: for the one or two songs whose words matter, get word timings and mark a window per song; no speech starts within 2 s before it or overlaps it; the envelope is held at 0 dB through it.

6. Levels: each song keeps its own weight

• Per-cue target −17 LUFS integrated for under and up before any envelope; music-up −16 to −14 short-term in the clear.
• Song-relative offset: offset = clamp(0.5 × (song_LUFS − playlist_median_LUFS), −2, +2); a ballad sits ~2 LU under a party track as it did on the night. Verify: up cues spread 3–5 LU across the film (not 0). Source: CAS 2026 (Becoming Led Zeppelin: one EQ/limiter across recordings and "some of that goes away").
• Nothing on the music bus but a panner: no afftdn, loudnorm, dynaudnorm or EQ on mx cues (a plan check greps the filter graphs). Provenance: the exact masters the subjects heard, not remasters.
• Live music is not speech. A processional, a band, a dance, an acoustic guitar on location never go through the speech chain (de-esser, compressor, voice EQ): highpass=f=40..60 only, target like up (−17 to −20), not ducked, excluded from the duck sidechain, and given a solo listen (music + live-music stems alone) before the full mix. Verify: those cues' rendered filter graphs contain no deesser or acompressor; crest factor within 1 dB of the source. Source: BAFTA Factual pattern; Summer of Soul (a separate music mixer).

7. Key lines and the ending

• Key line: from 4–6 s before, music fades (6 s) to −20 LU under dialogue or out; textures fade; bed stays. During: bed only, no spots, no music entry. After: 2–3 s of bed only (the breath), then the next element enters on a phrase boundary with a 2 s fade. Sources: Senna; Burns ("nothing makes sound more effective than the absence of sound"); Albrechtsen; Apollo 11 ("a few seconds of breathing room before the actual ignition").
• Laughter peaks: the recorded laugh is the sound; no laughter spot within 5 s; music holds its duck through it (keep-spans).
• Ending ladder: last line → ≥ 3 s bed only → bed fade 2 s → 1.5 s true silence → (optional: one recorded motif) → credits song from ss=0, fade_in=0, played to its real ending → 2 s silence → black. No music overlapping the final line; no spot effect in the last 30 s; nothing starts in the last 1.5 s before the credits; anything still sounding fades before it (music ≤ 4 s, dialogue within its air, ambience/live ≤ 2 s). The music budget skips the last narrated chapter so its song can swell as the button. Verify: silencedetect=noise=-60dB:d=1.2 finds exactly one silence between the last line and the credits song; the last non-silent event is the motif or the song's final note. Sources: Third Coast (Finn and the Bell: "listen to the very last second"; Time Bandit: ending music is editorial).
• A recurring motif, if any, is a recorded sound from the film's own sources, used 3–4 times ≥ 20 min apart; a generated stinger reads as a stinger.

8. Verify

See references/song-maps-and-posts.md for beat/section tooling and the post-planning algorithm.