Dev notes ·

Demo sounds, first pass with Stable Audio Open (local)

  • The demo's art is complete, but none of the 80 demo rows in audio-tracker.md had been made; the game runs on synthesised placeholders and free packs (three pack licences still marked "confirm" in the credits).
  • Made 114 sounds locally with Stable Audio Open in ComfyUI (2 takes each, 228 files, ~13 min on the RX 7900 XT): 6 ambience beds, 5 human-layer stand-ins, the core SFX split into one file per sound (e.g. footsteps per ground, damage per type), 7 UI writing-implement sounds, 3 stingers, 8 boss tells, 7 creatures × idle/attack/hurt/death, 11 Hoplite skills.
  • Masters: WAV 48 kHz 24-bit; one-shots trimmed and peak -1 dBFS; loops cross-faded so they wrap (ambience ~43 s, the model's limit is 47 s).
  • Saved to audio-work/embers-of-olympus/_tests/2026-10-26-stable-audio/ with index.html to listen and pick takes. Not in Unity, tracker unchanged until the user picks.
  • Left out: the 7 music tracks (composer or a music model), engine effects (SF-12, SF-15, SG-06), voice-only rows (SF-17, SF-20, SF-28, HM-03, HM-04).
  • Picked a take for each of the 114 demo sounds with CLAP (laion/larger_clap_general, Apache 2.0, in ComfyUI-Shared/models/clap/) plus checks for clipping, hiss and dead air: 94 clear picks, 15 close calls, 5 marked "redo" (UI_InventoryMoveEquip, UI_TabSwitch, the dog whine, the loot-beam hum, the poison hiss). Claude can't hear, so these are scored, not heard; the user's ears decide. Picks are pre-ticked in _tests/2026-10-26-stable-audio/index.html and listed in PICKS.md.
  • Redid the 5 weak sounds (new prompts, 4 takes each, rescored; all now above the weak line), then put all 114 picks in audio-work/embers-of-olympus/{02-ambience,03-sfx,04-ui,05-stingers}/ under the tracker file names (FINAL_PICKS.md in the _tests/2026-10-26-stable-audio folder lists the take behind each).
  • Wired into Unity (Assets/_Project/Audio/Demo/: Ambience 3, BossTells 7, Creatures 24, Skills 11; the rest stay on the NAS only):
    • Ambience: Rarian Plain → the Explore cue, Eleusis → the Hub, Ploutonion → the Underworld (replaces the synthesised Amb_Fields/Town/Cave; the cut-grove bed has no cue yet).
    • Polydegmon's tells: Grasping Toll, Cold Exhale, Intake and Deepening Pools (the pool-opening sound), The Received (both), The Pull; the generic BossTell and BossPull sounds. The Warden's tells stay synthesised.
    • New hooks: EnemyDefinition idle/attack/hurt/death sound lists (EnemyBrain plays them: attack at wind-up, hurt at most twice a second, one idle growl anywhere at a time within 30 m) for the Skeletal Hoplite, Keres, Shade, Field Rat, Dryad Blight and First Furrow (+ the scythe sweep on its attack); SkillDefinition.CastSounds (PlayerAudio plays them) for the 11 Hoplite skills.
    • Kept the licensed pack sounds where the game already used one (level-up, quest, waypoint, death stingers, shrine, portal, UI clicks, footsteps, hits): they are real recordings, and the AI stingers are the weakest of what Stable Audio makes. The music is still the pack tracks / synth placeholders.
    • Not wired (no hook yet): the other creature and boss sounds (Polydegmon's creature set, First Furrow's raise-the-dead), footsteps on six grounds, doors/chests/jars, the dog's sounds, loot pickups, the human layer, the cut-grove and hearth beds, UI writing sounds.
  • Tracker: 65 demo rows marked 🎧 (made), 27 also 📦 (in the engine); ledger AI-579..AI-643. The Stable Audio Open licence is the Stability AI Community License (free commercial use under US$1M annual revenue; may need registration at stability.ai/license): marked CHECK before release.

All dev notes