Skip to content

Sound sourcing — acceptance standard

Companion to sound.md (editorial) and sound-production.md (machine contract). Defines what makes a SOURCE file acceptable for an ambient track. Applied by the sourcing-research pass (docs/sound-library-sourcing-candidates.md) and every future wave. Canon: docs/superpowers/specs/2026-06-17-sound-library-sourcing-research-design.md.

Technical

  • Format: lossless preferred — WAV/FLAC, ≥16-bit (24-bit ideal). MP3 only where no lossless exists, ≥192 kbps. (Pipeline masters to AAC; lossless avoids double-compression.)
  • Sample rate: ≥44.1 kHz.
  • Channels: true stereo (2ch) for all nature/ambient — required for the planned spatial bake (naluma-app#687). Reject dual-mono masquerading as stereo. (Noise Colours excluded — omnidirectional.)
  • Duration: the pipeline loops the master seamlessly (equal-power crossfade), so playback length is unbounded — the constraint is loop fatigue (the ear locking onto a recurring signature event), not total length. Hard minimum > 2× loop crossfade (> 4 s at the 2000 ms default). Target by texture, and always prefer the longest clean candidate within a use_case:
  • Stationary beds (steady rain, fan, waterfall, steady stream, brown/white noise, room tone): ≥ 60–120 s suffices — no distinctive events to notice on repeat.
  • Eventful scenes (birdsong, crickets/frogs, breaking waves, thunder, struck bells/chimes): target ≥ 5 min, ideally 10–20 min, so no signature event recurs audibly on the loop. Never accept a short eventful clip just to fill a track slot — ship fewer tracks instead (depth of quality over slot count; see docs/sound-library-positioning.md).
  • Loopability: steady-state texture; no abrupt events or unique "signature" sounds near head/tail (a lone thunderclap/bird call recurring every loop is disqualifying for a masker). A long source is the cheapest insurance against this — events spread out, so the loop seam lands on neutral texture.
  • Clean signal: no clipping; low noise floor; no hum/handling noise; no human speech, no music, no recognizable melodies, no identifiable voices; no sudden level jumps; headroom for −16 LUFS.
  • Spectral fit to use_case: maskers need broadband energy; sleep sounds avoid harsh HF transients.

Editorial

  • Matches the track's value_prop/use_case (Gentle Rain = soft steady, not a downpour).
  • Scene purity: only the named sound, nothing unexpected (no traffic under a forest stream).
  • Authentic field recording (not obviously synthetic) unless the track is explicitly synth.
  • Warm brand register consistent with the album.

License

Priority order: 1. CC0 — Freesound license:"Creative Commons 0", OpenGameArt CC0. Preferred (no attribution). 2. CC-BY 4.0 — acceptable downgrade (commercial OK, attribution required — log it for an in-app credits screen). Also Pixabay Content License (no attribution; note Pixabay re-hosts as MP3, so confirm ≥192 kbps + true stereo on audition). 3. Paid royalty-free — effectively ruled out. The consumer royalty-free industry (Pond5, Envato, Artlist, Epidemic, Storyblocks, Boom Library, A Sound Effect, Epic Stock Media, Big Fish, Listening Earth) almost universally (a) forbids the end user extracting/downloading the audio file — which an offline-caching app does — and (b) requires the audio be synced with other media / not the "primary value", which an ambient-sound app violates. Only custom/enterprise licenses can fit (e.g. Epic Stock Media's Custom Application License, Boom Library's developer license) — a deliberate marquee-track decision, never a checkout purchase; get end-user-offline-caching + perpetuity confirmed in writing.

  • Must permit commercial use + embedding in a shipped app where the audio file is cached to the end user's device — this clause (not "commercial use") is what most licenses actually fail.
  • Free To Use Sounds is NOT clean for app use — its free tier excludes apps / "original sounds" and directs developers to a paid developer license. Do not treat it as permissive-free here.
  • Reject: CC-BY-NC, BBC RemArc, Zapsplat standard tier, ElevenLabs SFX (library use), per-sound-ambiguous.
  • Verify the license on the source page itself, not a search snippet — Freesound's search-listing license label is unreliable (it has shown CC0 / CC-BY-NC reversed); trust the sound page.
  • Capture provenance per file: source URL, author, license name + license URL, date checked.

Cost

  • Target €0 (CC0/permissive). Paid only to fill a genuine gap, ≤€30/track, flagged for approval.
  • No purchases or downloads during research.