ambience-and-sound-for-stills
Ambience beds, gap fill and spot effects for narrated documentaries, especially films built from photographs and silent archive: bed levels and coverage (never dead air), the gap lift, the layer hiera
SKILL.mdskills/ambience-and-sound-for-stills/SKILL.md
GitHubambience-and-sound-for-stills
A pause in a narrated film must sound like a place, not a dropout. A photograph gets a sound
only when something visible would make one, placed further away than the narrator. OpenDubStage
tool: opendubstage foley (opendubstage.foley: plan, check, Rules, levels,
prompts.lint, post.ERA_EQ, post.ROOMS).
1. Beds: never dead air
| Rule | Number | Source |
|---|---|---|
| A bed runs under every narrated stretch | per-cue target −40 LUFS at a −24 dialogue anchor (−16 LU); duck −3 dB under speech | Torcoli et al. 2019: ambience ≥ 15 LU under commentary |
| Not inaudible | −44 LUFS with a −5 duck (−49 under speech) is inaudible on a TV at living-room level; beds are the room the voice lives in, not a safety net against the silence detector | NPR §5 "what pros would say"; measured in practice |
| Gap lift | +5 dB over any stretch > 3 s with no speech and no music, eased 2 s; never above −33 LUFS short-term | measured in practice (45 accidental holes in one mix) |
| Gaps are air, not holes | momentary loudness (400 ms) in every speech gap ≥ −46 LUFS; ambience in the clear −38…−42 LUFS | Torcoli; EBU Tech 3341 |
| Dead-air detector | no span below −60 dBFS sample peak for > 3 s between first and last cue (strict: −50 dBFS / 2 s); programme head and tail exempt | opendubstage qc dead_air; ffmpeg silencedetect |
| Coverage check | the union of bed cues (target ≤ −38) covers [0, total] with no hole; dead air usually comes from beds shorter than their scenes, so check every bed's source length against the scene it must cover before mixing | measured in practice |
| Designed silence | at most two per film, chosen by hand, listed in an explicit whitelist that the dead-air check reads; inside one, nothing plays, not even the bed | Albrechtsen (total silence once or twice per film); Burns |
| Bed joins | successive beds crossfade 5 s (start 5 s early), never fade from silence at a chapter start; one bed per scene, chosen from the scene's pictures (time of day, place), not per still | foley.Rules.bed_crossfade_s; TSNGO (distance per scene) |
| Loops | 3–5 s equal-power; lengths that are not multiples of each other (41 s / 67 s) | Pixflow; rule of thumb |
| Nature beds under voice | low-pass 8 kHz so pines and crickets sit behind the voice and do not read as "nature documentary" | measured in practice |
| Room-tone floor | a learned floor from the narration's own gaps at −46 LUFS under the whole dialogue stem (see dialogue-editing-and-restoration §3) |
NPR room-tone practice |
| Perspective | on a few hero stills only, ±1–2 dB of bed top end following a push-in or pull-out; nothing else follows the picture move | Milano on Apollo 11 ("sweep the treble and go with the motion") |
2. Layers and levels (relative to dialogue)
| Layer | What | Level vs dialogue | At −24 |
|---|---|---|---|
| Bed | continuous scene ambience | −16 LU | −40 |
| Texture | semi-continuous and placeable: distant voices, water, birds, a far train; 10–40 s, 3 s fades; one at a time, changes with the scene | −12 LU | −36 |
| Spot | one-shot matched to a visible source | −8 LU (hero still −4) | −32 (−28) |
| Designed transition | a door, a car passing, applause, a shutter that is the cut; hand-requested, ≤ 1 per chapter | −2 LU | −26 |
Nothing in the effects stem ever reaches dialogue level (levels.MAX_OFFSET_LU = −2).
Distance is the dial: the narrator is present; the photograph is a memory; a sound that
comes across nearer than the narrator is wrong (Heywood, TSNGO: distance, not level, keeps a
long film from becoming monotonous). Murch's two-and-a-half layers: dialogue + bed leaves half a
layer for spots and music together.
3. Density and timing (foley.Rules defaults)
| Rule | Value |
|---|---|
| spots per still | ≤ 1; only visible sources (Kwok, TSNGO) |
| spots per minute | ≤ 4 under narration; ≤ 6 in a narration-free montage; 0 under music-up; 0 under live or sync sound |
| spacing | ≥ 5 s between spots; ≥ 30 s before the same class repeats; the same sample never within 600 s; 3 variants per class, rotated |
| placement | cut + 0.8 s ± 0.3 s, never within 0.4 s of the cut (a sound on the cut reads as a stinger); if blocked, the still's midpoint once |
| end | 0.1 s before the next cut; a spot belongs to its picture |
| spot fades | in 0.03 s, out 0.3 s |
| narration guard | nothing in the 0.5 s before a narration onset, nor on its first 0.3 s |
| key lines | bed only from 4 s before to 3 s after; textures end before the window |
| textures | one at a time, 10–40 s, 3 s fades |
| designed transitions | ≤ 1 per chapter, by hand |
| picture moves | spots never follow the Ken Burns move; no whooshes, risers or stingers (the lint rejects them) |
Natural density is low: measured on a large set of captioned photographs, roughly one spot-worthy object per 5–6 stills. Do not invent sounds for most pictures.
4. When not to add sound
- Portraits, close-ups, posed pictures, smiles (a face ≥ 3 % of the frame counts as a portrait).
- Details and documents: cake, rings, dress, letters, maps, certificates.
- Technically weak pictures (blurry, dark, overexposed): the picture is already asking for forgiveness.
- Objects that make no sound: a lantern does not tick; a padlock being clicked shut does, only if the hand is in the picture.
- Anything with real sync sound (speeches, performances, interviews on camera): the camera already heard it (Apollo 11 left the EVA audio alone).
- 360° spheres and wide empty landscapes: bed plus one slow directional wind texture, no spots (Fire of Love: "winds give it a sense of direction").
- Inside a key line, and inside a designed silence.
5. Avoiding the cartoon feel
- One action, one material, one place, one distance per prompt; tail "no music, no voices" (Spiegel: a Panzer "in snow", "on wet road", "on dry road").
- Lint (
foley.prompts.lint) rejects whooshes, swooshes, risers, swells, braams, impact hits, stingers, comedic sounds, music, intelligible speech, "epic" or "cinematic" scale. - Match the era and the mic (
post.ERA_EQ):highpass=f=80,lowpass=f=9000,equalizer=f=3500:t=q:w=1.2:g=-3, 30 ms fades, then loudness-normalise to the layer table. Libraries and generators "sound too modern" (TSNGO; Burns' team); Albrechtsen records foley "badly" on purpose to match rough production sound. - Worldize (
post.ROOMS): a short room matched to the bed: trees / open land (early reflections only,0.4 s), street, small room (0.5 s), hall (~1.1 s). The narrator's own room is never applied to the photograph's sounds. - Prefer textures to hits: laughter as a distant group, not a sitcom laugh; a few steps on pine needles, not a march; one glass at a table, not a toast cascade.
- Keep source artefacts: handling noise, the maker's breath, interjections are texture (Cameraperson, Fire of Love).
- Witness test: play a viewer five stills with and without the effects. If they cannot say which is which but prefer "with", the level is right; if they can point at the effect, lower it 3 dB or remove it (Burns' veterans: "Yes, that's the way it was").
- Invent only from evidence, then age it (Apollo 11's 1202 alarm built from the NASA spec and dirtied).
- Textures and beds:
loop=true, 22–30 s, low prompt influence (0.3) for natural variation; spots 1–3 s at higher influence (0.6); three variants each; keep a manifest with prompt, parameters and date so generation is reproducible.
Prompt template and keyword → class map: references/prompts-and-classes.md.
6. Workflow with opendubstage foley
opendubstage foley analyze SEQUENCE.json [--vision] # scene analysis per still (keywords, or Apple Vision labels)
opendubstage foley plan SEQUENCE.json -o cuesheet.json [--vision] # analysis -> cue sheet under the rules; rejected candidates are listed in "decisions"
opendubstage foley check SEQUENCE.json cuesheet.json # re-verify a (hand-edited) cue sheet; exit 1 on any violation
opendubstage foley generate SEQUENCE.json cuesheet.json --backend library|woosh -o fx
opendubstage foley render cuesheet.json -o stems # per-layer stems with ffmpeg
plan always passes check (property-based tests assert it). Hand edits go through check.
Resolution order when choosing a cue: the picture's moment type wins (speeches, performances, portraits →
none); then the no-sound classes; then the first matching spot; then texture; the bed from the
scene's median time of day and the majority bed selector among its pictures.
7. Verify
| Check | Threshold |
|---|---|
| bed coverage | union of bed cues covers [0, total]; every bed source ≥ its scene length (or looped) |
| gaps | momentary ≥ −46 LUFS in every speech gap; no dead_air outside the designed-silence whitelist |
| effects stem | short-term never above −28 LUFS; spot count per 60 s window ≤ 4 under narration |
| placement | no spot within 0.4 s of a cut, within 0.5 s before a narration onset, or inside a key-line window |
| anti-cartoon | every generated file passed the lint, ERA_EQ and a room; the witness test was run on ≥ 5 stills |
| textures | ≤ 1 at a time; 10–40 s |