Minara Video Music & SFX Generator

/minara-music — Generate music + sound effects from the script

Turn the video's per-second script into a music cue sheet, cues.json, then synthesize the BGM and picture-synced sound effects locally. Synthesis is pure local numpy and the output is deterministic (the same cues + seed always produce the same audio). It calls no external service and costs nothing.

Source: the scoring approach from the Minara V2 launch film, generalized into a reusable script; examples/v2-launch.json is that film's cue sheet.

Workflow (in order)

1. Read the script: get the per-second script from minara-video (or the storyboard / voiceover timing sheet). If the script has "sound atmosphere" or "music" columns, follow them; if not, judge from the mood of the picture.
2. Write the cue sheet: in the project's audio-src/cues.json, write:
   • duration (= final video length), bpm (default 120), key (default A minor; for bright/celebratory pieces you can use C major), seed.
   • sections: cut the script into music sections, pick one energy per section, and use note to describe the matching picture (for easy cross-reference).
   • sfx: each visible action in the script (tap, typing, order filled, notification popping up, transition, lock landing) gets one sound effect, timed to the second given in the script.
3. Show the operator the sheet first; don't render yet: in markdown tables, list "time range / energy / matching picture" and "sfx time / type / matching action", and wait for them to say OK before rendering. They edit the sheet, not the code.
4. Render:

   python3 ~/.claude/skills/minara-music/scripts/make_music.py <project>/audio-src/cues.json -o <project>/assets [--vo <project>/assets/vo.wav]
   

5. Self-check: the script prints each section; then measure loudness per section to confirm climax is loudest, breakdown/outro clearly drop back, and ducking is working. For listening issues (too busy, a sound effect stealing focus), adjust level / gain and re-render; each render takes a few seconds.
6. Deliver: give the path to bgm.wav (or bgm-duck.wav), and use open -R to highlight preview-mix.wav so they can listen. In HyperFrames, add it as an audio track, with the voiceover on a separate track.

Output

File	Contents
bgm.wav	Music + sound effects, 48 kHz stereo, -1 dBFS peak, 0.4 s fade-in at the start and 3.2 s fade-out at the end
bgm-duck.wav	Generated when there is a voiceover: music is lowered by duck_db (default -8 dB) where speech occurs; sound effects are not ducked
preview-mix.wav	Generated when there is a voiceover: music + voiceover mixed together, for listening only, not for the final video

energy section types

energy	Sound	Used for
silent	No music	Pauses that need complete silence
intro	Low drone + soft pad	Opening, logo, data converging
light	Pad + light arpeggio + sparse beat	Mobile interactions, lightweight feature demos
build	Adds bass and a kick on every beat, automatic riser at the end of the section	Gradual build-up leading to a key moment
full	Full drum kit + clap + bass + arpeggio	Continuous showcase of core features (backtesting, going live, automation)
breakdown	Drops drums and bass, leaving only pad and a soft arpeggio; the kick returns in the last bar	Scene changes, sections about security/privacy that need to "breathe"
climax	Fullest: full drum kit + lead melody + crash at the section start; automatically adds a riser if the previous section is not build	Product name / launch claim appears
outro	Resolves on the tonic chord + brand bell, fading out with the whole mix	End card, CTA

Each section can take "level": 0.5–1.3 to fine-tune its volume. Chords follow global bars (minor i–VI–III–VII, major I–V–vi–IV), so they don't break across sections.

Sound effect types (sfx)

type	Sound	Typical action	Optional parameters
tap	Short click	Button tap, selection	gain
typing	Continuous typing ticks	Typing in an input field	chars, cps (characters/second, matching the on-screen typing speed)
success	Two rising bell tones	Order filled, strategy running	gain
notify	Three-note alert	Notification or reminder popping up	gain
bell	Single bell tone	Avatar/logo appears	note (MIDI, default 81)
blips	Short tones stepping upward	Backtest progress, data loading	count, step (seconds), note
sparkle	Shimmering grains + breathy noise	Particles converging, opening	dur
riser	Rising noise	Manually added transition build-up	dur
whoosh	Reverse sweep	Window collapsing, push/pull transition	dur
thud	Low lock-in thud	Lock, confirmation, heavy landing	gain
crash	Cymbal	Extra accent	gain

Every sound effect can take gain to adjust its volume.

Rules

• Sound effects follow visible actions only: if there is no matching action on screen, add no sound effect, the same as "don't add text casually".
• Copy timings from the script; don't estimate: the sfx t uses the actual seconds from the per-second script / animation timeline; cps for typing uses the typing speed in the animation.
• If the script's length or order changes, update cues.json first and re-render; don't stretch the audio by hand.
• Don't use external AI music services or tracks downloaded from the web; when commercially licensed music is needed, the operator provides the file.