DigDeeper · experiments›all projects

decks.de → YouTube / SoundCloud matching

decks.de has no embeddable player, so every decks track needs a matched YouTube or SoundCloud source to be playable. This is the matcher running over a random sample of the archived decks backlog — play the matches and judge them yourself. Click why? under any match to see how it flowed through the matcher (which query won, what noise was stripped, how the artist was confirmed).

74% of decks tracks get a playable match (YouTube or SoundCloud). YouTube carries most of it (72%); SoundCloud rescues 2 tracks YouTube missed. The remaining 26% have no playable source and would be parked as no_match_found — most are decks rows with junk titles, not matcher failures.

72%YouTube match (68/95)

39%SoundCloud match (37/95)

74%playable (either source)

+2tracks SoundCloud rescued from a YouTube miss

26%no match → no_match_found

Confidence colour: ≥0.90 strong 0.80–0.90 good 0.75–0.80 borderline. Threshold to accept is 0.75. No ground truth here — colour reflects the matcher's own score, not verified correctness.

#	decks track	YouTube	SoundCloud

How it works

For each track the matcher builds several query variants (cleaned artist - title → title-only → raw → aggressive de-noised), runs a yt-dlp ytsearch/scsearch, and scores every candidate on fuzzy title+artist similarity, with boosts for a matching channel and penalties for conflicting mix types (Dub vs Original). First variant over 0.75 wins.
Various-Artists fix: decks frequently sets artist="Various Artists" and hides the real artist in the title (eric borgo - freaky). The matcher now splits that out — this alone lifted those tracks from ~0.51 (rejected) to ~0.85 (accepted).

Where it can improve

decks metadata noise is the main blocker. Of the 25 no-match tracks, 2 have a generic/positional title (a1, original, mix1) with no real song name — nothing to search on. Fixing DecksDownloader's per-track title extraction would recover more of these than any matcher tuning.
5 of 100 tracks failed metadata fetch entirely (fetch_metadata ValueError) — a separate decks-downloader bug; those can't be analysed at all.
1 matches landed in the 0.75–0.80 borderline band — most worth a manual listen; a few are likely false positives where the title words coincide. Tightening the artist-presence guard could trim these.
SoundCloud is a fallback, not a primary. Most SC hits duplicate a YouTube hit; it only earns its place by rescuing the YouTube misses — so the pipeline should try YouTube first, SoundCloud only on a miss.

Sample: 100 random archived decks tracks (5 excluded for metadata-fetch errors). Generator: digdeeper-similarity/scripts/build_decks_match_experiment.py. Lab notes — confidence is the matcher's own score, not verified ground truth; decks audio is a 60 s preview.