I’m the creator of a small game-audio store and wanted to share the approach behind my latest pack, Haunted Atmosphere.
Most horror ambiences are a single fixed track. I split each of 6 scenes into 4 stems:
- bed: drone / room tone that is always there
- air: wind, hum, vents, water
- detail: creaks, drips, crows, chains
- presence: high tension cluster and wordless whisper-like textures
Played together they reproduce the full mix sample for sample, so in-engine you can start all four and fade layers in as the player gets closer to the threat, then fire a stinger. Every loop is 60 s and seamless by construction (synthesized as a loop, no crossfade), with loop points embedded in WAV and OGG.
I’m the creator. Sounds are synthesized by code written with AI assistance; no generative AI audio or image models were used. AI tool: Grok Bot coding assistant (used only to write the synthesis code; current version as of Oct 2026). (Techniques: filtered noise, modal resonators, creak models, formant filters, synthetic reverb.)
Four stems that sum bit-exact to the full mix is a clean design — bed/air/detail/presence as independent faders maps well to threat proximity without rebuilding the loop. Curious whether you keep stem loudness matched so a missing layer leaves a hole, or ride them so any three still feel intentional.
I’m the creator, good question. They’re ridden, not matched. Each stem gets its own fixed level in the scene code. Then one shared gain brings the full mix to -20 LUFS, so the stems still sum to the mix exactly. Measured with ffmpeg ebur128 across the 6 scenes:
- bed: -24.9 to -22.9 LUFS
- air: -24.7 to -22.4
- detail: -32.5 to -28.0
- presence: -29.4 to -26.8
- full mixes: -20.0, true peak -8.5 to -6.1 dBTP
So bed and air are the loudest layers, and detail/presence sit 4-8 dB lower by design. (The synthesis code was written with AI assistance.)