Audio Stems

Audio Stems

Produce a piece's audio as independent stems — voiceover.mp3, music.mp3, sfx.mp3 — placed on a known timeline, plus an optional ducked master.mp3. Delivering stems separately lets you (or the user) iterate on audio without touching the rendered video, and drop tracks onto any editor.

Generate (ElevenLabs — scripts/elevenlabs.py)

• Music bed: python3 scripts/elevenlabs.py music "upbeat but understated corporate tech, instrumental, ~120 BPM" music.mp3 --ms 150000 (caps ~150s; for longer, acrossfade-loop). Keep it understated so VO leads.
• SFX palette: python3 scripts/elevenlabs.py sfx "soft cinematic whoosh" sfx_clips/whoosh.mp3 --seconds 0.9 — build a small palette (e.g. whoosh/swish, pop, type, generate, shimmer, ding, logo, crack). Min duration 0.5s. Always check level (ffmpeg -af volumedetect); regenerate if max , "events": [ {"t":12.34,"clip":"whoosh","gain":0.6}, … ] } (one silent base + each clip delayed to its time), then:

python3 scripts/build_sfx.py sfx_events.json sfx_clips sfx.mp3

Place events at known choreography moments — derive them from a scene model or from vo-sync's timing.json (t = scene.start + cue.at) so a sound lands exactly on the spoken word / the card raise. Same tool builds the VO track if you put per-scene clips in a folder named by id.

Tasteful placement (learned defaults)

• One transition sound (swish/whoosh) per push/scene-change; a softer/low one is less fatiguing than a bright hissy whoosh — offer a few options and let the user pick.
• pop on each card raise; type while text types, generate during a "processing/generating" beat; shimmer on a decrypt/sparkle; ding on success; logo sting on intro/outro; crack synced to a fracture.
• Keep per-interaction SFX quiet (gains ~0.3–0.5); the SFX track should be mostly silence.
• Don't put events in the 0→first-motion intro; verify every t ≤ total.

Mix (optional master — scripts/mix.sh)

./scripts/mix.sh music.mp3 voiceover.mp3 sfx.mp3 master.mp3

Side-chains the music under the voiceover (ducks when the voice is present), sums VO + SFX, and limits. Pad/trim every stem to the exact timeline length first (afade/apad). Target mean ≈ −20 dB, peak < −1 dB.

Deliver

Hand over the stems separately by default (voiceover.mp3, music.mp3, sfx.mp3) — only mux into the video when the user confirms. To mux onto a silent render: ffmpeg -i video.mp4 -i master.mp3 -c:v copy -c:a aac -b:a 192k -shortest out.mp4.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: JagZ999
• Source: JagZ999/explainer-video
• License: MIT

Install and usage instructions live in the source repository linked above.