Add sound effects and ambience

Two kinds of sound, handled differently:

• Spot effects: one event on picture (a door creaks open, a cup is set down, a title pops). Cue to the frame.
• Ambience: a bed under a whole scene (birds in a garden, wind, city hum, room tone). Loops and crossfades, sits low, changes only at cuts.

If hn4a offers ElevenLabs' sound-effects model, realistic effects come from text (elevenlabs-sound-effects, which checks). Otherwise pick a route per sound:

Ask which route when realism matters and it's unclear; don't present synthesized birds or doors as real recordings. Route B: put the sound in that model's prompt grammar (load its …-prompt-format leaf) and keep the shot's audio when splicing.

Place sounds against picture

• Cue spot effects from the picture: a 6 fps contact sheet of the moment, then frame-step. Transients land about 10 ms before the visual contact. For composited graphics, read cues from the compose spec instead (python-synth-sfx-for-motion-graphics).
• Pan toward where the source is in frame. Equal-power pan: L*=sqrt((1-p)/2), R*=sqrt((1+p)/2); for movement, ramp p in 10 ms slices.
• Ambience: trim a seamless loop (crossfade the loop point over 0.5-1 s), run it under the whole scene, fade in/out over 0.3-1 s at scene edges, and change it only at cuts.
• Levels: dialogue first. Ambience about 18-25 dB under dialogue peaks; spot effects under or level with dialogue. Duck ambience 6-10 dB under speech if it fights.
• Mix: build the track in Python at 48 kHz stereo, normalize the peak to about -1 dBFS, write audio/sfx.wav, and either list it in a compose spec's audio (with the original sound as another entry at lower volume if it stays) or mux with ffmpeg (-map 0:v -map 1:a -c:v copy -c:a aac). Final loudness for delivery is in guide-prepare-renders-for-delivery.

Route C recipes

Seed the noise generator so rebuilds match; butter(..., output='sos') + sosfilt.

• Pop, whoosh, squash, splat, thwack: python-synth-sfx-for-motion-graphics.
• Birdsong (stylized): a bird is a phrase of 3-7 chirps; a chirp is a sine whose pitch glides between 2-6 kHz over 40-120 ms (f0+(f1-f0)*(t/d)**0.7, optional 35 Hz trill of a few hundred Hz), shaped by sin(pi*t/d)**2, with 20-80 ms gaps. Schedule phrases at random (exponential gaps, ~1 per second), each with its own base pitch, pan in ±0.8 and gain 0.2-0.6. Add a short reverb or a little band-passed noise for air. Keep it well under dialogue.
• Door creak (stick-slip): an impulse train whose rate glides (e.g. 40 to 140 Hz over 1-1.5 s) with ~25% wobble and jitter, exciting four narrow band-passes (~420, 960, 1750, 3100 Hz at gains 1, 0.7, 0.45, 0.2). Slower glide for a heavy door, higher bands for a small hinge. Add a low thud (60-90 Hz decaying) if it closes.
• Wind: noise band-passed 150-900 Hz, amplitude driven by noise low-passed under 2 Hz (gusts). Rain: dense random clicks (band-passed 2-8 kHz) over pink-ish noise. Room tone: brown noise low-passed at 400 Hz, very quiet.
• Things synthesis does badly: voices, crowds, music, footsteps on specific surfaces, anything the audience knows the real sound of up close. Use route A or B for those.

Check

• You can't hear. Measure what you can (volumedetect on a silent span; peaks per cue) and put the result on a review page with each sound listed at its time in the row's body, and send the MP4 into the chat as well. Ask the person to listen; adjust gain, timing or the recipe from their notes and rebuild from the script (scripts/make_sfx.py), never by hand.
• If the person wants real recordings and has none, say plainly that hn4a can't generate realistic effects yet and offer the routes above.