Fair Comparison: DCCE vs. Strong Baselines (supervision-tiered)
All numbers are macro-F1 on the single fixed seed-42 split — the split DCCE's encoder was trained on, so its test set (277 codas) is genuinely held out. This is the only leakage-free basis for comparing a trained model (DCCE) against frozen baselines, so it is what we use for the head-to-head.
Why not the 10-split error bars for DCCE? They are valid for the frozen baselines (WhAM/AVES/BirdNET never saw our codas), but not for DCCE: DCCE was trained on the seed-42 train split, so reshuffling leaks its training codas into the probe's test set and inflates its scores (e.g. unit jumps 0.878 → 0.962). We therefore report DCCE on the held-out seed-42 split only.
Individual ID (the headline task) — tiered by supervision
Tier A — label-free representations (apples-to-apples)
Encoder never saw individual-ID labels. This is the fair test of our contribution.
| Model | Individual ID (macro-F1) |
|---|---|
| Raw Mel | 0.280 |
| BirdNET | 0.322 |
| AVES | 0.418 |
| WhAM | 0.426 |
| Raw ICI | 0.493 |
DCCE (no L_id) |
0.571 |
Tier B — ID-supervised representations
Encoder trained with individual-ID labels.
| Model | Individual ID (macro-F1) |
|---|---|
DCCE-full (with L_id) |
0.834 |
All three tasks, single-split seed-42
| Model | Tier | Unit | Coda Type | Individual ID |
|---|---|---|---|---|
| Raw ICI | floor | 0.599 | 0.931 | 0.493 |
| Raw Mel | floor | 0.740 | 0.097 | 0.280 |
| WhAM (best layer) | label-free | 0.895 | 0.261 | 0.426 |
| AVES (best layer) | label-free | 0.905 | 0.290 | 0.418 |
| BirdNET | label-free | 0.863 | 0.272 | 0.322 |
DCCE (no L_id) |
label-free | 0.883 | 0.673 | 0.571 |
DCCE-full (with L_id) |
ID-supervised | 0.878 | 0.578 | 0.834 |
What this says (honest interpretation)
- The contribution survives the fair test. Even with no individual-ID supervision, DCCE's architecture reaches 0.571 on individual ID — higher than every off-the-shelf label-free encoder (WhAM 0.426, AVES 0.418, BirdNET 0.322) and than raw ICI (0.493). So the dual-channel design itself carries individual-ID signal that generic and whale-specific encoders do not. (+0.145 over WhAM, label-free.)
- Supervision is roughly half the original headline. ID supervision lifts DCCE from 0.571 → 0.834. The old "0.834 vs WhAM 0.454 = +83%" framing conflated architecture and supervision; the honest decomposition is: +0.145 from the architecture (fair tier) and a further +0.263 from ID supervision.
- Raw ICI is a surprisingly strong individual-ID baseline (0.493) — it beats all three learned audio encoders. Honest and worth stating: rhythm timing carries more individual signal than any frozen audio embedding.
- Coda type: raw ICI (0.931) dominates every learned model, confirming coda type is essentially the ICI pattern (so no stronger rhythm baseline is warranted).
- Removing
L_idslightly improves DCCE on coda type (0.578 → 0.673) — with no individual-ID head competing for capacity, the rhythm/L_typesignal is cleaner.
Suggested honest claim for the paper
Among label-free representations, DCCE's dual-channel architecture attains the best individual-ID macro-F1 (0.571), exceeding the whale-specific generative model WhAM (0.426) and bioacoustic transfer encoders AVES (0.418) and BirdNET (0.322) on the same held-out split. Adding individual-ID supervision raises this to 0.834. The advantage over prior representations is therefore not an artifact of supervision: the architecture alone improves individual identification, and supervision compounds it.