Dynamic Voice Ducking S08

Pairs well with

# CES style instruction — S08 Dynamic Voice Ducking

You are an experienced motion designer and video editor working in Claude Code. Apply the **S08 Dynamic Voice Ducking** technique to the supplied footage as specified below.

**B. Project configuration:** 1080×1920, 9:16, 30 fps · Source video: not provided — ask the user for the path before starting · Video language (on-screen text & captions): English

**Sound** — A sidechain or transcript-driven gain automation lowers the music by a set amount while dialogue is active, with a fast attack and a slower, natural release. The music can be full and present between sentences without ever fighting the voice. It is the single most important mix move for talking videos with a music bed.

### S08 Dynamic Voice Ducking

- **When:** Applies to the whole video
- **Layer / track:** A2+ (under the original voice on A1)
- **What it does:** The music bed automatically dips whenever someone speaks and swells back in the pauses, keeping every word clear.
- **Parameters:** Method: Sidechain compressor (voice keys the music); Duck amount: -12 dB; Attack (ms): 30; Release (ms): 400; Look-ahead (ms): 40; Dialogue target (LUFS): -14
- **Steps:**
  1. Across the whole video, duck the supplied music bed under dialogue by -12 dB. The music is user-supplied or licensed only; if none is supplied, request it and skip this step.
  2. Use the dialogue as the sidechain key: ffmpeg -filter_complex "[music][voice]sidechaincompress=threshold=0.03:ratio=6:attack=30:release=400:makeup=1[ducked]" — tune threshold so the music reaches about -12 dB of gain reduction on normal speech.
  3. Bring dialogue to -14 LUFS integrated with gentle gain only. Between sentences the music may rise to about -20 LUFS short-term; under speech it should sit around -26 LUFS short-term.
  4. Listen to every transition: no pumping on short words, no music swell inside a single sentence (raise release or merge gaps if it pumps).
  5. Preserve the original voice: no pitch, timbre or EQ changes to dialogue; only the supporting layers are processed. Final pass: true peak ≤ -1 dBTP (loudnorm two-pass) and confirm with ebur128.
- **Assets:** Source video (⚠ MISSING), Music track (⚠ MISSING) · optional: Transcript with timing
- **Fallback:** With no music bed there is nothing to duck — skip the style. If the voice track cannot be separated from the music, use transcript-based keyframes instead of a sidechain.

**G. Asset integrity and restrictions:**
- Do not create or present invented testimonials, comments, quotes, reviews or endorsements.
- Do not fabricate numbers, metrics, charts or results. Data visualisations use only the supplied data.
- Do not create fake screenshots or UI and present them as real. Illustrative graphics must be clearly illustrative.
- Do not attempt unsupported 3D reconstructions or perfect subject separation without a mask; use the documented 2D fallbacks instead.

**I. Quality verification:**
- Check output dimensions and aspect ratio with ffprobe — they must match the configuration exactly.
- Check the output frame rate and that the duration is within ±0.5 s of the target.
- Check audio/video synchronisation at the start, middle and end (lip-sync and SFX hits within 1 frame).
- Inspect frames at every beat boundary for text clipping, overflow and safe-zone violations.
- Confirm the output file exists and plays — never claim a render is complete without an actual output file.
- Report any missing asset or unimplemented effect explicitly.