Literature Foundations for the Dual-Channel Contrastive Encoder (DCCE)

Project: Beyond WhAM — Dual-Channel Contrastive Encoding for Sperm Whale Coda Understanding Purpose of this document: a citation-ready evidence base for the Related Work / Background of the paper. It is organised as a chain of evidence that justifies the DCCE design, answering three questions in order:

1. What is known about sperm whale codas? (the biological substrate — why two channels, why individual ID is meaningful)
2. What is known about identifying individual animals from sound, across all taxa? (the task is real, recoverable, and increasingly done with learned embeddings)
3. What ML methods make DCCE possible, and where did each design choice come from? (the technical lineage)
4. Synthesis — every DCCE design decision mapped to the evidence above.

Conventions - Each bullet = one paper, in the form: Authors (Year) — Title — venue / ID → key finding → DCCE tie-in. - All arXiv IDs / DOIs were verified against primary sources where possible. Items whose metadata could not be fully verified are marked ⚠ verify and listed again in the appendix. - Three papers were read in full text for this review and yielded corrections to the previous draft's claims — these are flagged 🔧 and collected at the end.

PART 1 — Sperm Whale Codas: The Biological Substrate

1.1 What codas are, and the identity hierarchy

• Watkins & Schevill (1977) — Sperm whale codas — J. Acoust. Soc. Am. 62(6):1485–1490 (DOI 10.1121/1.381678). First scientific description of codas as patterned, repeated click sequences used socially rather than for echolocation. → DCCE tie-in: establishes the coda as the fundamental unit of analysis — the object DCCE encodes.
• Rendell & Whitehead (2003) — Vocal clans in sperm whales — Proc. R. Soc. B 270:225–231 (DOI 10.1098/rspb.2002.2267). Defines the vocal clan: large groups sharing a coda dialect, socially (not genetically) transmitted, maintaining separation even in geographic overlap. → DCCE tie-in: the top of the identity hierarchy; grounds the "clan → unit → individual" task structure and the claim that coda style is learned (hence learnable by a model).
• Gero, Whitehead & Rendell (2016) — Individual, unit and vocal clan level identity cues in sperm whale codas — R. Soc. Open Sci. 3(1):150372 (DOI 10.1098/rsos.150372). Read in full. Shows codas carry identity at three nested levels, but each level lives in different acoustic structure: the stereotyped 1+1+3 type is a clan marker shared by all whales (classifies individuals at near-chance, 33%); 4-click types discriminate units (88% via discriminant analysis, chance 54%); and individual identity is recoverable only within a unit, via fine variation in the 5R1 coda type (73–83% accuracy). Dataset: 1,502 codas attributed to 18 individuals across 9 units, 2005–2010. → DCCE tie-in: this is the single most important biological justification for the whole paper. (a) It proves individual identity is genuinely encoded in coda acoustics — so a learned embedding can recover it. (b) It proves individual ID is hard and local (within-unit, one coda type) — explaining why it is the task where all baselines struggle and where a purpose-built encoder should help most. (c) The 1,502-coda / 18-individual set maps directly onto the DSWP range our labels are drawn from. 🔧 Correction: the previous draft implied individual ID is broadly recoverable; in fact it is reliable only within a unit and primarily on 5R1.

1.2 The two-channel structure: rhythm (timing) vs. spectral (vowel-like) information

• Beguš, Sprouse, Leban, Silva & Gero (2025) — Vowel- and Diphthong-Like Spectral Patterns in Sperm Whale Codas — Open Mind (MIT Press) 9:1849–1874 (DOI 10.1162/OPMI.a.252). Read in full. Discovers that the spectral content of clicks is discretely structured into vowel-like categories — an a-coda (single formant ~5,780 Hz) and an i-coda (two formants ~3,757 / 6,579 Hz) — plus diphthong-like formant trajectories, analysed via a source–filter model. 1,214 focal codas / 7,181 clicks / 10 individuals. → DCCE tie-in: this is the empirical basis for DCCE's spectral channel. It shows that within-click spectral texture is a real, discrete, independent information axis — exactly what the CNN spectral encoder is meant to capture, and why a mel-spectrogram (not just ICI timing) is necessary.
• Beguš, Dąbkowski, Sprouse, Gruber & Gero (2026) — The phonology of sperm whale coda vowels — Proc. R. Soc. B 293(2069):20252994 (DOI 10.1098/rspb.2025.2994; bioRxiv 2025.06.09.658556). Argues these spectral "vowels" not only sound vowel-like but pattern like linguistic vowels, forming a dimension that interacts with (does not replace) the rhythm/timing system; introduces a short-i vs. long-ī duration contrast. → DCCE tie-in: justifies treating rhythm and spectral content as two interacting channels rather than one signal — the core architectural premise of DCCE. 🔧 Correction: "Beguš et al." is two papers (Open Mind 2025 + Proc. B 2026), not one; the vowel inventory is a/i + diphthongs, not a large set.
• Leitão, Lucas, Poetto, Hersh, Gero, Gruber, Bronstein & Petri (2023/2025) — Evidence of Social Learning Across Symbolic Cultural Barriers in Sperm Whales — arXiv:2307.05304. Models rhythmic micro-variation within codas using Variable-Length Markov Chains on discretised ICI sequences; shows non-identity codas converge in style across clans (social learning) while identity codas stay distinct. Explicitly notes its model uses temporal rhythm only — spectral information is absent. → DCCE tie-in: (a) demonstrates the rhythm channel carries unit-/clan-level signal at the micro-variation level (motivating a sequence encoder, not a bag of ICIs); (b) its stated limitation — "spectral information absent" — is precisely the gap DCCE's second channel fills.
• Hersh, Gero, Rendell, Whitehead, et al. (2022) — Evidence from sperm whale clans of symbolic marking in non-human cultures — PNAS 119(37):e2201692119 (DOI 10.1073/pnas.2201692119). Defines identity codas and shows clans actively maintain distinct markers. → DCCE tie-in: explains why some codas resist convergence and others don't — i.e., why rhythm alone is an unreliable identity cue and a multi-signal representation is needed.

1.3 Combinatorial and contextual structure

• Sharma, Gero, Payne, Gruber, Rus, Torralba & Andreas (2024) — Contextual and combinatorial structure in sperm whale vocalisations — Nature Communications 15:3617 (DOI 10.1038/s41467-024-47221-8). Read in full. Decomposes codas into four combinatorial features — rhythm (18 duration-normalised types), tempo (5 types), rubato (smooth modulation across consecutive codas), and ornamentation (extra clicks, ~4% of codas) — yielding at least 143 frequently-realised combinations and roughly doubling per-coda information capacity (~5 → ~10 bits). Releases the 8,719-coda DominicaCodas dataset (EC-1 clan, ~60 whales, 2005–2018). → DCCE tie-in: (a) source of the label taxonomy and the DominicaCodas alignment our dswp_labels.csv is built from; (b) shows coda structure is rich and partly contextual, supporting a learned representation over hand-crafted features. 🔧 Correction: the real claim is "≥143 frequently-realised combinations" and an information doubling, NOT a "vocabulary ~10× larger." Fix this number in the draft.

1.4 Machine learning on sperm whale codas (the immediate prior art)

• Bermant, Bronstein, Wood, Gero & Gruber (2019) — Deep Machine Learning Techniques for the Detection and Classification of Sperm Whale Bioacoustics — Scientific Reports 9:12588 (DOI 10.1038/s41598-019-48909-4). Read in full. First deep-learning pipeline for codas: CNN click detector (99.5%) + LSTM/GRU coda-type classifier (97.5% over 23 types), and individual identification — but only between two whales (99.4%). → DCCE tie-in: the closest prior deep-learning attempt at sperm whale individual ID. DCCE's 12-individual linear-probe task is a far harder, more realistic setting — this paper is the natural "first attempt" to contrast against, and its use of a GRU for ICI sequences directly precedes DCCE's rhythm encoder.
• Gubnitsky, Avidor Mevorach, Gero, Gruber & Diamant (2024) — Automatic Detection and Annotation of Sperm Whale Codas — arXiv:2407.17119. A signal-processing + graph-clustering pipeline that detects codas and separates them from echolocation using multipulse structure and resonant-frequency constraints; discovers new coda types. → DCCE tie-in: the upstream detection/annotation step our labelled corpus depends on; also the tool that would be needed to extract true ICIs from generated audio (a limitation in our synthetic-augmentation experiment).
• García, Seetharaman, Kumar & Pardo (2023) — VampNet: Music Generation via Masked Acoustic Token Modeling — arXiv:2307.04686 (ISMIR 2023). Masked-token transformer over a neural audio codec for music generation. → DCCE tie-in: the backbone WhAM fine-tunes; understanding it clarifies that WhAM is a generation-first model whose representations are a by-product.
• Paradise, Sharma, Muralikrishnan, Chen, Flores García, Pardo, Diamant, Gruber, Gero & Goldwasser (2025) — WhAM: Towards a Translative Model of Sperm Whale Vocalization — arXiv:2512.02206 (NeurIPS 2025). Fine-tunes VampNet on ~10,000 codas with a masked-acoustic-token objective; evaluates the emergent representations on rhythm / social-unit / vowel classification (linear probe). → DCCE tie-in: the primary comparison target. It is explicitly generative, so its classification ability is emergent, not designed — the gap DCCE addresses by being a representation-first, identity-targeted encoder. It is also trained on ~6.7× our data, framing the "domain prior vs. scale" question.
• Goldwasser, Gruber, Kalai & Paradise (2023) — A Theory of Unsupervised Translation Motivated by Understanding Animal Communication — arXiv:2211.11081 (NeurIPS 2023). Proves error rates in unsupervised translation fall as the source language's complexity rises, so translating a sufficiently complex animal system is feasible. → DCCE tie-in: theoretical motivation for building good representations of coda structure first; supports a bottom-up, structure-respecting approach.

PART 2 — Acoustic Individual Identification of Animals (the task, across taxa)

This is the centrepiece: evidence that "identify the individual from its vocalisation" is a real, established, and increasingly machine-learnable problem — the exact task DCCE targets.

2.1 Individual vocal identity is real and behaviourally used

• Carlson, Kelly & Couzin (2020) — Individual vocal recognition across taxa: a review of the literature and a look into the future — Phil. Trans. R. Soc. B 375(1802):20190479 (DOI 10.1098/rstb.2019.0479). Comprehensive cross-taxa review of individual vocal recognition; documents how widespread it is and identifies gaps (taxonomic bias, temporal stability, specificity). → DCCE tie-in: the canonical citation establishing that vocal individuality is a general animal-kingdom phenomenon — DCCE's individual-ID task is not a sperm-whale curiosity but an instance of a broad, well-studied problem.
• Janik, Sayigh & Wells (2006) — Signature whistle shape conveys identity information to bottlenose dolphins — PNAS 103(21):8293–8297 (DOI 10.1073/pnas.0509918103). Playback experiments show dolphins extract identity from the frequency-modulation contour of signature whistles even after voice/timbre cues are removed — identity lives in the learned pattern shape, not just the voice. → DCCE tie-in: a strong cetacean precedent that individual identity is carried by structured acoustic pattern — exactly what a sequence-/shape-aware encoder is built to capture.
• King & Janik (2013) — Bottlenose dolphins can use learned vocal labels to address each other — PNAS 110(32):13216–13221 (DOI 10.1073/pnas.1304459110). Wild dolphins respond selectively to copies of their own signature whistle — evidence the signal functions as a referential, individually distinctive "name." → DCCE tie-in: reinforces that cetacean vocalisations carry robust, learnable individual signatures, so recovering them with a classifier is biologically well-posed.
• Nousek, Slater, Wang & Miller (2006) — The influence of social affiliation on individual vocal signatures of northern resident killer whales — Biology Letters 2(4):481–484 (DOI 10.1098/rsbl.2006.0517). Killer-whale calls carry individual signatures whose distinctiveness depends on social proximity: across matrilines calls are highly distinctive, but within a matriline they converge and individual signatures weaken. → DCCE tie-in: a near-perfect parallel to sperm whales — individual signatures nested within social structure, with within-group convergence. This explains why individual ID within a unit is the hard case and motivates DCCE jointly modelling unit and individual identity.

2.2 Machine-learning / deep-embedding approaches to individual ID

• Nolasco, Moummad, Stowell & Benetos (2024) — Acoustic identification of individual animals with hierarchical contrastive learning — arXiv:2409.08673 (⚠ verify ICASSP 2025 acceptance — cite as preprint). Frames acoustic individual identification (AIID) as a hierarchical, multi-label problem and proposes hierarchy-aware contrastive losses so embeddings preserve species/taxa structure while separating individuals; supports open-set (novel-individual) scenarios. → DCCE tie-in: the most methodologically parallel paper to DCCE. It is contrastive learning for individual ID with an explicit identity hierarchy — directly echoing DCCE's clan→unit→individual structure and validating the contrastive-embedding approach to AIID. This should be cited prominently.
• Tolkova, Klinck, Cusano, Kügler & Parks (2025) — Detection, communication, and individual identification with deep audio embeddings: a case study with North Atlantic right whales — bioRxiv 10.1101/2025.07.11.664307 (preprint). Uses BirdNET embeddings + a shallow classifier to identify individual right whales from 234 upcalls across 11 individuals at 3 sites, and studies noise robustness. → DCCE tie-in: a near-identical regime to ours — learned embeddings + linear evaluation for cetacean individual ID on a small dataset (~11 individuals, few hundred calls). The strongest "this exact protocol works for whales" precedent for DCCE's setup.
• Sarkar & Magimai-Doss (2023) — Can Self-Supervised Neural Representations Pre-Trained on Human Speech distinguish Animal Callers? — Interspeech 2023, arXiv:2305.14035. Probes 11 human-speech SSL models on marmoset calls and finds their embedding spaces carry individual-caller information recoverable without fine-tuning. → DCCE tie-in: supports the core thesis that learned (even cross-domain) representations encode individual identity recoverable by a simple probe — and motivates including human-speech SSL models (wav2vec2/HuBERT) as baselines.
• Trapanotto, Nanni, Brahnam & Guo (2022) — Convolutional Neural Networks for the Identification of African Lions from Individual Vocalizations — Journal of Imaging 8(4):96 (DOI 10.3390/jimaging8040096). Transfer-learned CNNs (VGG16/ResNet50/AlexNet) identify individual lions from roars at >98% on a tiny dataset (164 samples, 5 individuals); compares spectrogram representations. → DCCE tie-in: a clean demonstration that CNN spectrogram embeddings do individual ID on very small data — direct support for DCCE's CNN spectral encoder in a low-resource regime.
• Guerrero et al. (2023) — Acoustic animal identification using unsupervised learning — Methods in Ecology and Evolution (⚠ verify — paywalled; author list / volume / DOI unconfirmed). Reportedly clusters/identifies individuals from vocalisations without labels. → DCCE tie-in: represents the unsupervised branch of AIID — a useful contrast to DCCE's structure-aware (unit/ID-supervised) contrastive approach. Verify before citing.

2.3 Benchmarks and the gap DCCE sits in

• Hagiwara, Hoffman, Liu, Cusimano, Effenberger & Zacarian (2023) — BEANS: The Benchmark of Animal Sounds — ICASSP 2023, arXiv:2210.12300. Standardised bioacoustics benchmark over 12 datasets, defining classification and detection task families (not a dedicated individual-ID task). → DCCE tie-in: the reference point for standardised bioacoustic evaluation — and evidence that no widely-adopted benchmark covers sperm-whale coda individual ID, motivating our own linear-probe / macro-F1 protocol. 🔧 Correction: contrary to some secondary summaries, BEANS does not include an individual-ID task.
• Hamer, Triantafillou, van Merriënboer, Kahl, Klinck, Denton & Dumoulin (2023) — BIRB: A Generalization Benchmark for Information Retrieval in Bioacoustics — arXiv:2312.07439. A retrieval-oriented generalisation benchmark stressing out-of-distribution retrieval, few-shot novel classes, and label shift, evaluated by representation learning + nearest-centroid search. → DCCE tie-in: defines the embedding-retrieval evaluation of frozen representations under distribution shift — a more demanding protocol our embeddings could be assessed against (future work), and a model for thinking about the recording-year shift.

PART 3 — Representation-Learning Methods: The Technical Lineage of DCCE

3.1 Self-supervised audio representation learning (the backbone paradigm)

• Baevski, Zhou, Mohamed & Auli (2020) — wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations — NeurIPS 2020, arXiv:2006.11477. Masks latent speech and solves a contrastive task over learned quantised units; CNN feature extractor + Transformer context net. → DCCE tie-in: the canonical SSL audio backbone and the "pretrain unlabeled → probe frozen" paradigm DCCE operates in; a natural off-the-shelf baseline encoder.
• Hsu, Bolte, Tsai, Lakhotia, Salakhutdinov & Mohamed (2021) — HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units — IEEE/ACM TASLP 29 (DOI 10.1109/TASLP.2021.3122291), arXiv:2106.07447. Masked prediction of clustered "hidden units." → DCCE tie-in: the masked-prediction alternative to contrastive SSL, and the architecture AVES (our bioacoustic baseline) is built on — so it doubles as the conceptual link between generic-speech and animal-sound encoders.
• Chen et al. (2022) — WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing — IEEE JSTSP 2022, arXiv:2110.13900. Extends HuBERT with denoising/overlap simulation; SOTA on the SUPERB probing benchmark. → DCCE tie-in: a strong frozen-feature backbone evaluated by exactly the linear-probe convention DCCE uses (SUPERB).
• Gong, Chung & Glass (2021) — AST: Audio Spectrogram Transformer — Interspeech 2021, arXiv:2104.01778. Convolution-free patch Transformer applied directly to log-mel spectrograms. → DCCE tie-in: situates DCCE's mel-spectrogram spectral encoder within the standard spectrogram-model family.
• Gong, Lai, Chung & Glass (2022) — SSAST: Self-Supervised Audio Spectrogram Transformer — AAAI 2022, arXiv:2110.09784. Self-supervised masked-patch pretraining for the AST. → DCCE tie-in: a directly comparable self-supervised spectrogram encoder; an alternative design point to DCCE's CNN.
• Chen et al. (2023) — BEATs: Audio Pre-Training with Acoustic Tokenizers — ICML 2023, arXiv:2212.09058. Iteratively co-trains an acoustic tokenizer and a masked-audio SSL encoder; SOTA general-audio classification. → DCCE tie-in: a leading general-audio SSL backbone — a candidate strong frozen-feature baseline for representation quality on bioacoustic transfer.
• Niizumi, Takeuchi, Ohishi, Harada & Kashino (2021) — BYOL for Audio (BYOL-A) — IJCNN 2021, arXiv:2103.06695 (journal: arXiv:2204.07402). Negative-free self-distillation on augmentations of single audio segments. → DCCE tie-in: positions DCCE in the contrastive-vs-distillation design space; a non-contrastive SSL alternative.
• Saeed, Grangier & Zeghidour (2021) — Contrastive Learning of General-Purpose Audio Representations (COLA) — ICASSP 2021, arXiv:2010.10915. Lightweight contrastive SSL — same-recording segments are positives — evaluated by linear probe across 9 tasks including animal sounds. → DCCE tie-in: a close conceptual sibling: contrastive audio SSL + linear-probe transfer that already includes animal-sound transfer; a strong methodological analogue for DCCE.

3.2 Contrastive and cross-modal learning (the direct ancestors of DCCE's objective)

• Chen, Kornblith, Norouzi & Hinton (2020) — A Simple Framework for Contrastive Learning of Visual Representations (SimCLR) — ICML 2020, arXiv:2002.05709. Introduces NT-Xent loss with same-sample augmented positive pairs; shows a projection head and temperature/L2-norm are essential, and pre-projection features probe best. → DCCE tie-in: the direct source of DCCE's loss (NT-Xent, τ=0.07), its projection/fusion head, and the "probe pre-projection embeddings" choice.
• Radford et al. (2021) — Learning Transferable Visual Models From Natural Language Supervision (CLIP) — ICML 2021, arXiv:2103.00020. Contrastive alignment of two modalities (image, text) via a symmetric loss over an N×N similarity matrix; positives are matched cross-modal pairs. → DCCE tie-in: the conceptual ancestor of DCCE's cross-channel positive pairs — DCCE replaces (image, text) with (rhythm, spectral) of same-unit codas.
• Elizalde, Deshmukh, Al Ismail & Wang (2023) — CLAP: Learning Audio Concepts From Natural Language Supervision — ICASSP 2023 (⚠ verify venue), arXiv:2206.04769. Dual audio/text encoders into a shared space for zero-shot audio classification. → DCCE tie-in: the audio specialisation of the CLIP idea; the bridge from CLIP to bioacoustic contrastive models (BioLingual).
• Wu, Chen, Zhang, Hui, Berg-Kirkpatrick & Dubnov (2023) — Large-scale Contrastive Language-Audio Pretraining (LAION-CLAP) — ICASSP 2023, arXiv:2211.06687. Open large-scale CLAP with feature fusion; SOTA text-audio retrieval. → DCCE tie-in: the open CLAP backbone underpinning later bioacoustic CLAP models; cite alongside Elizalde to disambiguate the two distinct "CLAP" works.

3.3 Speaker-embedding / metric-learning lineage (DCCE is a speaker-ID-style method — currently uncited)

This entire subsection is the literature the original draft omitted; DCCE's individual-ID embedding is conceptually a speaker-verification model.

• Dehak, Kenny, Dehak, Dumouchel & Ouellet (2011) — Front-End Factor Analysis for Speaker Verification (i-vectors) — IEEE TASLP 19(4):788–798 (DOI 10.1109/TASL.2010.2064307). Compresses a variable-length utterance into a single fixed "total-variability" identity vector. → DCCE tie-in: the original "utterance → fixed identity vector" idea DCCE modernises.
• Variani, Lei, McDermott, Lopez-Moreno & Gonzalez-Dominguez (2014) — Deep Neural Networks for Small Footprint Text-Dependent Speaker Verification (d-vectors) — ICASSP 2014 (DOI 10.1109/ICASSP.2014.6854363). First neural speaker embedding: classify speakers, then pool a hidden layer into a "d-vector." → DCCE tie-in: establishes the frame-encoder + pooling → identity-vector template DCCE's encoders follow.
• Snyder, Garcia-Romero, Sell, Povey & Khudanpur (2018) — X-Vectors: Robust DNN Embeddings for Speaker Recognition — ICASSP 2018 (DOI 10.1109/ICASSP.2018.8461375). TDNN + statistics pooling, trained discriminatively with heavy augmentation. → DCCE tie-in: the canonical supervised speaker embedding — the standard reference any new identity-embedding method (DCCE included) is positioned against.
• Desplanques, Thienpondt & Demuynck (2020) — ECAPA-TDNN — Interspeech 2020, arXiv:2005.07143. Adds Res2Net/SE blocks, multi-layer aggregation, and attentive statistics pooling; the de facto modern speaker-embedding backbone. → DCCE tie-in: the current SOTA speaker-embedding architecture — the human-domain analogue of what DCCE does for whales.
• Wan, Wang, Papir & Lopez-Moreno (2018) — Generalized End-to-End Loss for Speaker Verification (GE2E) — ICASSP 2018, arXiv:1710.10467. Per-batch loss pulling each embedding to its speaker centroid and pushing from others. → DCCE tie-in: a foundational centroid-based metric-learning objective for identity embeddings — the conceptual sibling of DCCE's contrastive identity loss.
• Chung, Huh, Mun, et al. (2020) — In Defence of Metric Learning for Speaker Recognition — Interspeech 2020, arXiv:2003.11982. Shows metric/contrastive losses (incl. an angular-prototypical loss) match or beat classification losses for open-set speaker recognition. → DCCE tie-in: directly justifies preferring contrastive/metric learning over plain classification when target identities are unseen — exactly DCCE's setting.
• Xiang, Wang, Huang, Qian & Yu (2019) — Margin Matters: Towards More Discriminative DNN Embeddings for Speaker Recognition — APSIPA ASC 2019, arXiv:1906.07317 (root loss: Deng et al., ArcFace, CVPR 2019, arXiv:1801.07698). Angular-margin losses sharpen identity embeddings (25–30% relative EER reduction). → DCCE tie-in: the supervised-margin counterpart to DCCE's contrastive separation of identities; a candidate enhancement for future work.
• Huh, Heo, Kang, Watanabe & Chung (2020) — Augmentation Adversarial Training for Self-Supervised Speaker Recognition — NeurIPS 2020 SAS Workshop, arXiv:2007.12085. Contrastive speaker embedding without labels, with an adversarial objective making it invariant to channel/recording cues while staying speaker-discriminative. → DCCE tie-in: a close analogue — label-free contrastive identity learning that explicitly disentangles identity from recording/channel confounds, the same confound DCCE's design and our recording-year analysis target.
• Zhang, Zou & Wang (2021) — Contrastive Self-Supervised Learning for Text-Independent Speaker Verification — ICASSP 2021 (DOI 10.1109/ICASSP39728.2021.9413351). SimCLR-style InfoNCE for speaker embeddings plus a channel-invariant loss to stop the network encoding session/channel info. → DCCE tie-in: explicit SimCLR-for-identity with channel disentanglement — the closest human-speech parallel to DCCE's contrastive dual-channel design and our recording-year confound concern.
• Cho, Villalba, Moro-Velazquez & Dehak (2022) — Non-Contrastive Self-Supervised Learning of Utterance-Level Speech Representations (DINO for speech) — Interspeech/J-STSP 2022, arXiv:2208.05413 (⚠ verify journal volume). DINO self-distillation learns utterance embeddings that transfer strongly to speaker verification, sometimes beating supervised x-vectors. → DCCE tie-in: the non-contrastive route to label-free identity embeddings — a design alternative that frames DCCE's contrastive choice.
• Nagrani, Chung & Zisserman (2017) — VoxCeleb: A Large-Scale Speaker Identification Dataset — Interspeech 2017, arXiv:1706.08612. The standard in-the-wild speaker-ID benchmark underpinning all the above. → DCCE tie-in: cite when stating DCCE tackles the whale analogue of the speaker-identification task.

3.4 Bioacoustic foundation models (domain-specific encoders to benchmark against)

• Hagiwara (2023) — AVES: Animal Vocalization Encoder Based on Self-Supervision — ICASSP 2023, arXiv:2210.14493. A HuBERT-base self-supervised encoder pretrained on curated animal/natural audio; strong transfer across bioacoustic tasks (incl. BEANS). → DCCE tie-in: the single most important off-the-shelf baseline — a self-supervised encoder built for animal sound, evaluated by probing. WhAM used it as a baseline; DCCE should too (we extract AVES embeddings on our exact split).
• Kahl, Wood, Eibl & Klinck (2021) — BirdNET: A Deep Learning Solution for Avian Diversity Monitoring — Ecological Informatics 61:101236 (DOI 10.1016/j.ecoinf.2021.101236). Widely deployed supervised CNN over bird spectrograms; its embeddings are routinely reused for transfer (incl. the right-whale individual-ID study above). → DCCE tie-in: the flagship supervised bioacoustic feature extractor; a domain baseline contrasting with DCCE's self-supervised contrastive approach.
• Ghani, Denton, Kahl, Klinck, et al. (2023) — Global Birdsong Embeddings Enable Superior Transfer Learning for Bioacoustic Classification (Perch) — Scientific Reports 13:22876 (DOI 10.1038/s41598-023-49989-z), arXiv:2307.06292. Shows embeddings from a large supervised bird classifier give strong few-shot transfer, beating general-audio embeddings. → DCCE tie-in: direct evidence for the frozen-embedding + shallow-classifier protocol DCCE uses, and a strong bioacoustic transfer baseline.
• Robinson et al. (2024) — Transferable Models for Bioacoustics with Human Language Supervision (BioLingual) — ICASSP 2024 (⚠ verify venue), arXiv:2308.04978. A CLAP-style language-audio model trained on >1M animal audio-caption pairs; zero-shot bioacoustic classification across 1000+ species. → DCCE tie-in: the bioacoustic specialisation of CLAP — a recent domain foundation model and an example of language-supervised vs. DCCE's self-supervised approach.
• Robinson et al. (2024/2025) — NatureLM-audio: An Audio-Language Foundation Model for Bioacoustics — arXiv:2411.07186 (⚠ verify ICLR 2025). First audio-language LLM for bioacoustics; SOTA zero-shot incl. unseen species; demonstrates speech/music → animal transfer. → DCCE tie-in: the most recent large bioacoustic foundation model and evidence for the cross-domain transfer premise underlying WhAM (music→whale) — useful framing for why scale-based generalists exist and where a small purpose-built model still wins.

3.5 Probing methodology (how we and WhAM are evaluated)

• Pasad, Chou & Livescu (2021) — Layer-wise Analysis of a Self-Supervised Speech Representation Model — ASRU 2021, arXiv:2107.04734. Probes wav2vec 2.0 layer-by-layer, finding an autoencoder-style trajectory where acoustic→phonetic→word information peaks at successive depths. → DCCE tie-in: the canonical justification for layer-wise linear probing — directly supports our per-layer probe of WhAM (and of the SSL baselines), and the finding that "best layer" differs by task.
• Castellon, Donahue & Liang (2021) — Codified Audio Language Modeling Learns Useful Representations for MIR (JukeMIR) — ISMIR 2021, arXiv:2107.05677. Probes a frozen generative music model's internal layers with shallow classifiers; middle layers carry the richest content. → DCCE tie-in: the precedent for probing a generative model's representations — exactly what we do to WhAM — and the source of the "middle layers probe best" convention behind our layer-10 choice.
• Rauch et al. (2025) — Can Masked Autoencoders Also Listen to Birds? (Bird-MAE) — arXiv:2504.12880 (⚠ verify full author list). A domain MAE for birds; argues linear probing under-reads frozen bioacoustic features and proposes "prototypical probing" (up to +37% mAP). → DCCE tie-in: an important caveat to cite when justifying our linear-probe protocol — it bounds how much our linear-probe numbers may understate representation quality, and points to a stronger evaluation for future work.

PART 4 — Synthesis: How the Evidence Justifies Each DCCE Design Choice

This is the argument the Related Work section should build. Each DCCE decision is mapped to its supporting literature.

• Why two separate encoders (rhythm + spectral)? Because the biology says codas carry two interacting information channels: timing/rhythm (Gero 2016; Leitão 2025; Sharma 2024) and discrete vowel-like spectral structure (Beguš 2025 Open Mind; Beguš 2026 Proc. B). Leitão 2025 explicitly flags that rhythm-only models leave spectral information on the table — DCCE's second channel fills exactly that stated gap.
• Why a GRU for rhythm and a CNN for spectral content? Bermant (2019) already showed GRU/LSTM models suit ICI sequences and CNNs suit coda spectrograms for sperm whales; Trapanotto (2022) shows CNN spectrogram embeddings do individual ID on tiny datasets. The architecture is the established tool for each channel, not an arbitrary choice.
• Why individual ID is the headline task (and why it is hard)? Gero (2016) proves individual identity is encoded but recoverable only within a unit on specific coda types; Nousek (2006) shows within-group signature convergence in orcas. So individual ID is biologically real but structurally hard — the regime where a generalist model (WhAM) should struggle and a purpose-built encoder should help. Bermant (2019) only ever did 2-whale ID; DCCE's 12-individual task is a genuine step up.
• Why a contrastive cross-channel objective (the core novelty)? Cross-modal contrastive alignment (CLIP 2021) and NT-Xent (SimCLR 2020) provide the loss and the cross-view positive-pair idea; the speaker-verification literature shows contrastive/metric learning is the right objective for open-set identity (Chung 2020; GE2E 2018) and that contrastive identity learning can be made invariant to recording/channel confounds (Huh 2020; Zhang 2021). DCCE = (CLIP's cross-view idea) ⊗ (speaker-embedding's identity objective), specialised to the rhythm/spectral channels.
• Why this is a representation-learning contribution, not "just another classifier"? WhAM (2025) is generation-first; its classification ability is emergent (probed à la JukeMIR 2021 / Pasad 2021). AVES (2023), BirdNET/Perch, BioLingual, NatureLM-audio are the bioacoustic-foundation-model context — large, general, scale-driven. DCCE's claim ("a domain-structured prior can beat scale on fine-grained identity in low-resource settings") is meaningful precisely against this backdrop.
• Why the small-data, frozen-embedding + linear-probe setup is legitimate? Tolkova (2025) does cetacean individual ID with frozen embeddings + shallow classifier on ~11 individuals; Sarkar (2023) recovers caller identity from frozen SSL features; COLA (2021) and Perch (2023) validate linear-probe transfer (incl. animal sound). Bird-MAE (2025) is the honest caveat that linear probing may understate representation quality.
• Why the recording-year confound matters (and is novel here). The speaker-SSL literature treats channel/session disentanglement as a first-class problem (Huh 2020; Zhang 2021) and BIRB (2023) benchmarks distribution shift — yet WhAM's evaluation did not control for it. Our Cramér's-V analysis imports a standard speech-domain concern into the coda setting.
• Why translate from human speaker recognition at all? Carlson (2020) and Janik (2006/2013) establish that individual identity in animal calls is a structured, learnable signal directly analogous to human speaker identity; the i-vector→x-vector→ECAPA→contrastive-SSL lineage is the mature toolkit for that exact problem. DCCE is the principled port of that toolkit to sperm whale codas.

APPENDIX A — Citations flagged for verification before final use

APPENDIX B — Three corrections to the previous draft (read papers)

1. Gero et al. (2016) — Individual ID is reliable only within a unit, primarily on the 5R1 coda type (73–83%); the shared 1+1+3 type classifies individuals at near-chance (33%). Do not claim broad individual recoverability.
2. Sharma et al. (2024) — The headline is "≥143 frequently-realised combinations" and a doubling of per-coda information (~5→~10 bits), built from rhythm/tempo/rubato/ornamentation — not "a vocabulary ~10× larger."
3. Beguš et al. — This is two papers: Open Mind (2025, spectral/acoustic discovery) and Proc. R. Soc. B (2026, phonological patterning). Vowels are a / i + diphthongs, framed as a channel that interacts with (not replaces) the rhythm/timing system.

APPENDIX C — Suggested grouping for the paper's Related Work section

1. Sperm whale communication & the two-channel coda → Part 1 (lead with Gero 2016 + Beguš 2025/26 + Sharma 2024 + Leitão 2025).
2. Acoustic individual identification across animals → Part 2 (lead with Carlson 2020 review; anchor on Nolasco 2024 + Tolkova 2025 as the method/regime precedents; Bermant 2019 as the sperm-whale-specific predecessor).
3. Representation learning we build on → Part 3 (SimCLR/CLIP for the objective; the speaker-embedding lineage for the identity framing; AVES/BirdNET/Perch as bioacoustic baselines; Pasad/JukeMIR for probing).
4. The gap & our position → Part 4 synthesis (WhAM is generative; no model encodes the two-channel decomposition; no benchmark covers coda individual ID).