The 100 tracks the engine calls most similar to the seed, in its own order. Play any of them,
and compare how six embedding strategies score the same pair.
SEED
Le Nomad — Two Left Turns 129.0 bpm · key G#
They mostly agree on which tracks are similar (every score is a tight 0.8–0.9) but disagree a lot on order — e.g. full-track keeps only 5/10 of the baseline top-10. The scores below are coloured by how each strategy differs from the baseline —
greener = rates it more similar,
redder = less.
3.6%average score difference from the baseline, across all five strategies
0.031cosine spread of the whole top-99 — why a ~4% shift still reorders the list
5/10of the baseline top-10 kept by the closest strategy (full track)
baseline middle-60 s vector the live engine serves (defines this order) · full Mean-pool over the whole track · mid·c A different, centred 60 s window (not the production 25%-start one) · b+m+e Begin+Mid+End averaged into one vector · mean·c Mean of section-aligned cosines · max·c Max cosine over all 3×3 section pairs
How the six strategies differ
All six use the same Discogs-EffNet model — they differ only in which audio they embed and how they compare it. Percentages are the average score change vs baseline across 99 tracks.
Single span · one mean-pooled vector
Embed one chunk of audio, average its ~1-per-second patch embeddings into a single 1280-d vector, compare with cosine.
baseline · middle 60 s production
A 60 s window starting at 25 % of the track. This is what beta serves — it defines the order on this page.
full track -1.3%
Mean-pool over the whole track. Closest to baseline; slightly lower on average because intros, breakdowns and outros dilute the core groove.
mid 60 s · centred -4.5%
A different middle minute — centred at 50 % instead of starting at 25 %. Same idea as baseline, different seconds; the gap shows how sensitive the result is to which 60 s you grab.
Begin / Middle / End · three sections
Embed three 60 s sections (start, centre, end) separately, then combine them — either into one vector, or by comparing sections.
b+m+e average -1.9%
Average the three section vectors into one, then cosine. A whole-shape summary of the track.
mean cosine -6.6%
Compare sections pairwise (begin↔begin, mid↔mid, end↔end) and average. Strictest — every section has to line up, so it scores lowest.
max cosine -2.4%
The single best-matching section pair across all 3×3 — “does any part of A sound like any part of B?” Most lenient.
What this shows. Similarity = cosine on the 1280-d Discogs-EffNet embedding.
The baseline column is what beta serves today; the others re-embed the same audio differently
(full track, a different 60 s, or begin/middle/end sections). Sample: 99 bandcamp neighbours of
one seed. Takeaway: the six strategies barely change whether a track is similar, but reorder the
list a lot — because everything here sits in a 0.03-wide cosine band. Read-only study; nothing written back.
Players stream from Bandcamp. 2026-05-29 · DigDeeper experiments.