Suggest Sfx

suggest-sfx — the SFX pass

Step 4 of the pipeline: take a video with its master cut + baked visual beats and add tasteful sound effects, synced to the visual beats and the narration, drawn from a shared, growing SFX library. Work collaboratively and beat-by-beat; the user audits the plan and gives notes.

Same shape as the rest of the pipeline: a declarative plan (sfx-plan.json) is the source of truth → a tool consumes it (tools/mix_sfx.py) → a library grows (media/library/sfx/) → a hard USER-AUDIT gate before anything is mixed.

Sound taste is a brand contract — read brand.md §10 every time. The house style is calm/premium (Linear / Anthropic / Vercel), felt-not-heard, always under the voice. NOT MrBeast-loud.

Inputs (read these first, every session)

• brand.md §10 (Sound design) — the SFX taste contract: subtlety, density, palette, levels, sync, signature motifs, source policy. Non-negotiable.
• videos/video-N/work/timeline.json — the shots and their master spans + type (cutaway/overlay). The cut-in/out boundaries are your transition (whoosh) candidates.
• videos/video-N/work/edited-transcript.json — word-level times (ms) in the MASTER timeline. Sync cues to these words (a pop on "boom", a chime on "free").
• The shot sources remotion/src/shots/video-N/*.tsx — the INTERNAL animation frames are where the real visual beats are (toggle flip, image reveal, page flip, badge pop). Read the shot, convert its local frame to master time: at_s = shot.master_in_s + local_frame / shot.fps. Do not guess — the cue must land on the exact frame the thing happens.
• media/library/sfx/catalog.json + palette.json — the library you draw from and grow.

The library (media/library/sfx/) — the durable asset

• media/library/sfx/palette.json — generation recipes: generic, reusable sound ids + prompts + duration + tags. The source of truth for what the library SHOULD contain.
• media/library/sfx/catalog.json — the manifest of what EXISTS: {id, file, category, tags, duration_s, peak_dbfs, loudness_lufs, source, model, license, prompt, used_in} per clip. Written by gen_sfx.py.
• media/library/sfx/clips/*.mp3 — loudness-normalized clips (~−20 LUFS, −1.5 dBFS ceiling) so a plan's per-cue gain_db is perceptually meaningful.
• Library-first, always. Reuse an existing clip before generating. Name/tag clips GENERICALLY (ui-toggle-on, whoosh-soft, pop-reveal) so future videos reuse them — the library is the durable asset, each video is one draw from it. Only genuinely-missing sounds get added to palette.json and generated. (SFX that must live INSIDE a Remotion shot instead go in media/library/sfx/ via staticFile().)

Workflow

1. Read brand §10 + timeline + transcript + the shot sources + catalog. Pull exact cue times: for each candidate beat, find the visual moment's local frame in the shot and/or the narration word in the transcript, and convert to master seconds.
2. Decide taste with the user up front if unsettled (present-ness, density, which kinds, source) — then apply brand §10. Don't re-ask settled decisions.
3. Propose the cue list — function-first. For each beat ask what it needs (brand §10): motion (whoosh) · tension (riser) · emphasis (impact/pop) · snap (click). Score key transitions, reveals, and tasteful click-sequences; a 3–4-shot click-sequence counts as ONE gesture. Mark genuinely deniable texture as "optional": true. Layer the 2–3 biggest moments (build-and-drop): riser→impact on a scripted reveal, whoosh→pop so a cut stands out — in the plan a layer is two events at the same/ adjacent at_s that sum. Sync each cue to the VISUAL beat and often the exact word.
4. Library-first sourcing. For each distinct sound the plan needs, reuse a catalog clip if one fits. For misses, add a generic recipe to palette.json and generate: python tools/gen_sfx.py (ElevenLabs Sound Effects API → normalized clip → catalog). Fall back to curated royalty-free only where generation is weak; record source + license either way.
5. Author videos/video-N/work/sfx-plan.json — one event per cue, at_s on the MASTER timeline. (Schema below.)
6. USER-AUDIT GATE (hard). Present the resolved cue sheet (python tools/mix_sfx.py <plan> --print) and get the user's approval/notes BEFORE mixing. This is the same gate as cuts.json / the edit plan.
7. Mix the audition preview once approved: python tools/mix_sfx.py videos/video-N/work/sfx-plan.json → videos/video-N/output/video-N-first60-sfx.mp4 (or the plan's render.out). Mix over any composited version with --preview <file>. Light sidechain duck under the voice + a safety limiter. Iterate on the user's notes (gains, timing, add/cut cues) — re-print, re-mix.
   • Verify audibility with numbers, not hope. After mixing, RMS-diff the mixed audio vs the voice-only preview at each cue window: a story-critical cue should add ≥ +4 dB, texture +1–3 dB. Inaudible knocks hide in the cue sheet alone — don't ship a cue you haven't confirmed lands.
   • Transient-clip gain gotcha. Short percussive clips (knock/stamp/keys/snap) hit the −1.5 dBFS peak ceiling BEFORE reaching the −20 LUFS loudness target, so they catalog ~3–5 dB quieter than sustained clips. Their plan gain_db must be ~3–5 dB higher than the brand §10 table implies (percussive knocks/ stamps landed at 0..+1). The audibility check above is what surfaces this.
   • Short-form calibration. Under a short's near-continuous narration (~2.7 w/s, few gaps) the brand §10 gain table is 4–8 dB too quiet across the board — the duck + voice masking eat everything. Short-form landed at: transitions/whooshes −3, story pops/impacts 0..+3, stamps/snaps 0..+7, layered-hero risers 0..+2. Start a short's plan there, not at the long-form table.
   • Cues that sit fully UNDER continuous speech never measure — accept felt-not-heard or cut them. Short-form click-sequences measured +0 dB at ANY gain (pushed to +9 experimentally: identical) — the sidechain duck suppresses quiet clips the whole time the voice is active. Set such cues at a sane +6-ish and let them live in the word gaps, or delete them; do NOT chase them with gain (a pause would make them spike).
   • Measure transients with tight windows. RMS over 0.6s dilutes a 50 ms click to nothing and an adjacent loud cue can leak in and fake a pass (a click "measured" +9.8 because a chime 0.17s later shared its window). Use ~0.3s windows for snap/pop/zap/stamp, and measure a riser at its final third (its energy is at the END).
   • Character beats level — a noisy TEXTURE cannot be fixed by gain (rejected twice in testing). Static/glitchy/hummy clips read as NOISE over speech even at −4 dB; lowering them just makes quiet noise. If the user says "noisy", swap the sound's CHARACTER (clean mechanical snap) or use silence — see brand §10 "no static/glitch textures under narration".
8. Save back. New clips stay in the library + catalog (they carry to the next video). Update used_in. A security/tech-themed set (knock-solid, glitch-zap, scan-hum, trap-snap, stamp-hit, keys-typing-soft) is library-first reusable for future videos.