Conference manuscript integration notes
Use final_paper.md as the conference-content baseline; it is LaTeX despite its extension. Keep the intended three-author conference header. The following replacement methods text describes the corrected experiment, not the legacy runs. Generated results tables are at results_v2/tables/main_results.tex; the human-readable result report is revision_results.md.
Replacement dataset and split text
“We use 1,501 DSWP audio files and retain 1,383 codas after excluding 118 noise-flagged files. These span three social units and 22 annotated coda types. Individual-identification experiments use 754 codas across 11 unambiguous, nonsingleton identity labels; 620 unknown-identity codas, eight codas with the unresolved label 6070/6068, and one singleton identity are excluded from identity losses and scoring, but retained for unit and type tasks. Dates are parsed explicitly in day/month/year format.
“We establish a shared train/validation/test assignment for every coda before representation training. Task-specific subsets inherit this assignment; representation training, positive-pair sampling, fitted preprocessing, and probe fitting use training codas only. We evaluate three ordinary stratified splits, each allocating approximately 64/16/20% to training/validation/test, and three distinct outer folds from a fixed five-fold partition grouped by recording date. Grouped validation also separates recording dates. Recording date is a proxy for recording session. Because some identities occur on only one date, we report the training/test class support and separately account for test identities absent from training.”
Replacement feature text
“Audio is converted to mono and resampled to 48 kHz. We use a 2,048-sample Hann STFT window, a 1,152-sample hop, 64 mel bands spanning 0–8 kHz, and per-clip peak-relative power in decibels with an 80 dB floor. Each example occupies 256 frames; shorter clips are padded at the floor. No clip in this corpus is truncated under this configuration. A fixed affine transform maps the dB range to [-1,1]. The temporal-mel feature floor concatenates means and standard deviations from four bins over valid frames. Nine-element annotated ICI vectors are zero-padded and standardized using training codas only.”
Replacement objective and control text
“The dual encoder comprises a two-layer GRU for annotated ICI and a three-block CNN for mel features. Their 64-dimensional outputs are concatenated and passed through LayerNorm, Linear(128,64), ReLU, and Linear(64,64). NT-Xent normalizes its inputs internally. The type auxiliary term is weighted cross-entropy on the rhythm embedding; the optional identity term is weighted cross-entropy on the spectral embedding, masked for ineligible identities. Both auxiliary coefficients are 0.5; class weights are computed from the training subset.
“We distinguish two pairing operations. The whole-coda condition pairs f(rA,sA) with f(rB,sB) for distinct training codas from the same social unit. The swapped condition pairs f(rA,sB) with f(rB,sA). Both use unit labels for pairing and unit-balanced sampling. Other same-unit examples may still be negatives under the pairwise NT-Xent objective. Consequently, neither condition is fully self-supervised. A no-ID condition removes only the identity auxiliary term.
“To control for task supervision, we train the same dual-input architecture with unit, type, and identity classification losses on its fused embedding. We also train shallow adapters on frozen final-layer AVES embeddings, with and without the annotated ICI input. These controls match downstream label access; the AVES backbone remains frozen. All neural conditions use 50 fixed epochs, AdamW at learning rate 0.001 and weight decay 0.0001, cosine scheduling, and batch size 64. We use the final epoch without test-dependent checkpoint selection.”
Replacement evaluation text
“For each task, a balanced logistic-regression probe (C=1, maximum 2,000 iterations) is fit to frozen representations using training codas and a training-fitted scaler. WhAM and AVES layers are selected using validation macro-F1; the selected layer is then evaluated on test codas. BirdNET uses its fixed cached embedding. Neural representations are retrained for every split/fold. We report per-run scores and mean ± sample standard deviation.
“Our primary macro-F1 averages over true test classes represented in training. We report eligible and supported sample counts and an additional score including unseen test labels. Candidate sets differ across grouped identity folds, so grouped and ordinary ID scores are not estimates of the same fixed-class task. Comparisons between models use identical folds and support. Historical scores are not directly comparable because this revision changes split membership, identity eligibility, preprocessing, and support conventions.”
Replacement recording-year interpretation
“Recording year and social unit are associated in the complete clean corpus (chi-squared = 530.04, df=6; Cramér’s V = 0.438). This motivates evaluation that separates recording dates. Association alone does not establish that a model uses recording artifacts, and the dual-channel design does not guarantee immunity to recording conditions. Validation-only WhAM layer profiles are treated as descriptive diagnostics, not independent causal evidence. The original year coverage and across-layer correlation are superseded by the corrected date analysis.”
Claims to remove or qualify when integrating results
- Remove the original 0.834/0.571 ID headline and any architecture-only decomposition of that difference. Use the corrected repeated results.
- Do not say no-ID DCCE is label-free or that frozen baselines received equivalent training supervision.
- Do not describe whole-coda pairing as channel swapping. Name the two conditions consistently in equations, diagrams, and tables.
- Replace “DCCE is structurally less exposed” with the actual grouped-split finding, including uncertainty and limits.
- Poor linear decoding does not establish loss of information during waveform tokenization; it establishes poor linear accessibility under the tested pooling/probe.
- Separate an architectural decomposition from empirical independence of biological channels. A mel spectrogram contains temporal structure, and raw ICI can carry identity cues.
- Narrow novelty to the exact sperm-whale-coda setting and acknowledge Tolkova’s multi-individual right-whale work.
- Limit generalization to this corpus and tested splits. These tasks do not establish semantic decoding of whale communication.
The bibliography corrections and source-build problems remain listed in reviewer_revision_audit.md. Before final submission, consolidate the conference source as a .tex file, resolve the architecture-image filename and conference style assets, verify every reference, insert generated figures/tables, and compile/render the conference PDF. The present work supplies corrected experiments and integration text; it is not a certified submission-ready PDF.