No single representation wins every task
Raw ICI attains 0.987 ± 0.009 random-split type macro-F1. ICI plus temporal mel attains 0.724 ± 0.039 individual-ID macro-F1. Strong timing references are essential when the type labels themselves reflect interval structure.
DCCE U+T attains mean random-split unit/type/ID scores of 0.901/0.865/0.597. Its fused-ID mean exceeds either separately probed branch (rhythm 0.459, spectrum 0.367), but remains below the simple combined-feature reference. With ID loss, DCCE U+T+I gives 0.907/0.735/0.581. In this configuration, identity supervision does not improve mean fused-identity performance.
Explore the main comparison
The chart shows the three tasks together. Switch the evaluation protocol or expand from four representative models to all 13 manuscript rows. Values and whiskers show mean ± sample SD across three runs. Model order stays fixed so the comparison does not imply a statistically established ranking.
Bar length: mean macro-F1. Whiskers and ± labels: sample SD. Range 0–1. These are not confidence intervals.
Primary scores use supported examples and classes. Grouped identity support changes across folds and differs from the random protocol. See [metric definitions and support counts](evaluation.html#what-macro-f1-measures-here) before interpreting the gap. ## The complete manuscript table| Representation | Random: unit | Random: type | Random: ID | Grouped: unit | Grouped: type | Grouped: ID |
|---|---|---|---|---|---|---|
| Raw ICI | 0.570 ± 0.023 | 0.987 ± 0.009 | 0.439 ± 0.007 | 0.540 ± 0.036 | 0.956 ± 0.036 | 0.379 ± 0.103 |
| Temporal mel | 0.773 ± 0.007 | 0.202 ± 0.025 | 0.429 ± 0.045 | 0.569 ± 0.094 | 0.069 ± 0.012 | 0.130 ± 0.074 |
| ICI + temporal mel | 0.837 ± 0.013 | 0.804 ± 0.016 | 0.724 ± 0.039 | 0.608 ± 0.150 | 0.668 ± 0.051 | 0.255 ± 0.062 |
| WhAM (frozen) | 0.906 ± 0.011 | 0.258 ± 0.046 | 0.450 ± 0.052 | 0.667 ± 0.155 | 0.066 ± 0.016 | 0.101 ± 0.075 |
| AVES (frozen) | 0.910 ± 0.010 | 0.253 ± 0.028 | 0.491 ± 0.029 | 0.700 ± 0.100 | 0.061 ± 0.022 | 0.074 ± 0.051 |
| BirdNET (frozen) | 0.887 ± 0.020 | 0.215 ± 0.053 | 0.480 ± 0.022 | 0.612 ± 0.133 | 0.063 ± 0.011 | 0.085 ± 0.022 |
| WhAM + ICI (frozen) | 0.921 ± 0.005 | 0.722 ± 0.027 | 0.598 ± 0.028 | 0.768 ± 0.057 | 0.376 ± 0.151 | 0.153 ± 0.135 |
| AVES + ICI (frozen) | 0.926 ± 0.011 | 0.667 ± 0.044 | 0.652 ± 0.044 | 0.712 ± 0.111 | 0.366 ± 0.084 | 0.128 ± 0.103 |
| BirdNET + ICI (frozen) | 0.884 ± 0.018 | 0.497 ± 0.028 | 0.580 ± 0.031 | 0.628 ± 0.095 | 0.287 ± 0.061 | 0.100 ± 0.055 |
| DCCE U+T | 0.901 ± 0.020 | 0.865 ± 0.115 | 0.597 ± 0.051 | 0.598 ± 0.078 | 0.692 ± 0.088 | 0.248 ± 0.130 |
| DCCE U+T+I | 0.907 ± 0.010 | 0.735 ± 0.091 | 0.581 ± 0.012 | 0.596 ± 0.052 | 0.628 ± 0.074 | 0.223 ± 0.175 |
| AVES FT | 0.924 ± 0.007 | 0.243 ± 0.025 | 0.414 ± 0.078 | 0.626 ± 0.157 | 0.072 ± 0.050 | 0.060 ± 0.026 |
| AVES + ICI FT | 0.913 ± 0.011 | 0.466 ± 0.059 | 0.544 ± 0.017 | 0.641 ± 0.158 | 0.211 ± 0.069 | 0.073 ± 0.038 |
All 78 mean/SD cells above come from the numerical-audit ledger for the final manuscript’s main table. The downloadable summaries also contain earlier corrected regimes, identified by their version field. This table selects the frozen/fixed references and the current uniform-sampling DCCE/adaptation rows used in the paper. Frozen-plus-ICI controls have their own summary file.
Input availability changes the comparison
Frozen AVES plus ICI attains random ID 0.652, compared with 0.491 for audio alone. Fine-tuned AVES plus ICI attains 0.544, compared with 0.414 without ICI. Annotated timing is additional input that a waveform-only model does not receive.
These controls improve the fairness of input comparisons without making training histories identical. Frozen models bring external pretraining, DCCE uses task-informed pairing and auxiliary objectives, and AVES adaptation has a different capacity and training budget. The results support comparing input and supervision conditions explicitly, rather than attributing every difference to the encoder architecture.
Branch performance does not guarantee fusion performance
An auxiliary identity objective improves the mean random spectral-branch ID score from 0.367 to 0.443. Over the same comparison, fused ID falls from 0.597 to 0.581. Rhythm-branch type remains 0.983/0.982 without/with ID loss, whereas fused type is 0.865/0.735.
Those are measurements of different outputs from the same trained checkpoints. They are not contradictory results. The ablation chapter separates branch diagnostics from independently trained architecture controls and explains why neither proves a biological division into independent channels.
What we can conclude
DCCE combines useful information, but gains depend on task, objective, input access, and partition. Simple combined features remain strong references, and these experiments do not establish that the fusion MLP is necessary or superior to joint concatenation. Three-run differences are descriptive and configuration-specific.
A broader claim about recognizing whales in new populations would require independent recordings and external-population evaluation. The next chapter examines how much performance changes even within this dataset when recording dates and years are held out.
Evidence: main summaries, frozen input controls, run-level predictions, and numerical audit.