It is the reference counterpart to lane-owned sample libraries like pop/hellsine/samples/, which hold AC's own recorded sounds. Sample audio here is third-party and stays local — only the metadata (per-source manifest.json + global INDEX.json + onsets.json) is committed, so anyone with the repo can rebuild the chops from the URL.

layout #

pop/samples/
  INDEX.json                       ← tracked · aggregates every source
  README.md                        ← this file
  <source-slug>/
    manifest.json                  ← tracked · slice recipe + chop list
    onsets.json                    ← tracked · librosa onset times (cheap)
    source.m4a                     ← gitignored · yt-dlp download
    source.wav                     ← gitignored · 48k mono for cutting
    chops/
      chop-001-0428ms.wav          ← gitignored · padded zero-index = ms length
      chop-002-1192ms.wav
      …

adding a source #

node pop/bin/sample-from-youtube.mjs <url> \
  [--slug my-name]        # override slug (default: from YT title)
  [--min-ms 250]          # minimum chop length
  [--max-ms 6000]         # maximum chop length (clamps long tails)
  [--max-chops 64]        # cap total chops, keep loudest
  [--start 12 --end 180]  # trim head/tail seconds from source
  [--note "for hippyhayzard verse 2"]
  [--redo]                # recut even if chops/ exists

Pipeline: yt-dlp → ffmpeg (48k mono PCM) → pop/bin/detect_onsets.py (librosa) → onset-aligned WAV chops with 3 ms declick fades + per-chop peak normalization → manifest + global-index update.

reproducing chops on a fresh checkout #

git pull
node pop/bin/sample-from-youtube.mjs <url-from-INDEX.json>
# or pass --slug to keep the existing directory name

yt-dlp is the only thing the recipe truly needs; onsets.json is kept around so the cutter is fully deterministic without re-running librosa if you only want to retune --min-ms / --max-ms (the cutter re-runs onsets every time today, but the data is committed for audit and downstream tools).

BBC Sound Effects (sound-effects.bbcrewind.co.uk) #

A second source: the BBC Sound Effects archive (30k+ effects). Fetched via pop/bin/bbc-fetch.mjs (lib: pop/lib/bbc-rewind.mjs) — no credentials needed, it hits the same public search + media endpoints the website's app uses, and writes full-quality WAVs named by BBC id (<id>.wav, e.g. pachinko-bbc/07022449.wav).

# search only — print results so you can pick ids:
node pop/bin/bbc-fetch.mjs --query "pachinko" --list

# search + download the top N into pop/samples/<slug>/:
node pop/bin/bbc-fetch.mjs --query "pachinko" --count 6 --slug pachinko-bbc

# download specific ids directly:
node pop/bin/bbc-fetch.mjs --slug pachinko-bbc --id 07022449,07032210

Each fetch writes a tracked manifest.json (id + description + the RemArc licence stamp) and drops a .gitignore so the WAV audio stays local, then upserts the global INDEX.json.

⚠ Licence — RemArc, non-commercial only. The BBC archive is free for personal / educational / research use and non-commercial live performance (e.g. the AC-native dj piece at a free/art set). It is not cleared for commercial release — a BBC sample must never be baked into a DistroKid release master. Every fetched sample is tagged license: "remarc-noncommercial", commercialUse: false. For commercial use the BBC licenses the same library via Pro Sound Effects. Terms: https://sound-effects.bbcrewind.co.uk/licensing

why third-party audio stays out of git #

Same posture as pop/references/README.md: third-party copyrighted material does not belong in a public github repo. The manifest + URL + slicing parameters are scaffolding (fair-use-ish for a private research notebook); the actual mp3/wav is not.