Baseline Approach for DCCE — Justification, Sources, and Plan

Beyond WhAM: Dual-Channel Contrastive Encoding for Sperm Whale Coda Understanding

Bottom line: We replace our current thin baseline set with strong baselines drawn from the bioacoustics literature — the same encoders that prior work uses for animal-sound and (specifically) cetacean individual-identification tasks — and we make the comparison against WhAM fair by controlling for label supervision. Every baseline below is justified either by a published paper or by an explicit logical argument; none is chosen arbitrarily.

1. What we mean by "baseline," and an honest diagnosis of the current draft

• A baseline is an external comparison method — a model that already exists, run on our data and metric, to test whether DCCE is actually better than the alternatives. (Standard ML evaluation methodology.)
• This is distinct from an ablation, which removes a component of our own model to test which parts matter. Our rhythm_only / spectral_only / late_fusion / full rows are ablations, not baselines. (Methodological distinction; conflating the two is part of why the current rationale reads unclear.)
• Diagnosis of the current table: of our listed "baselines," only WhAM is a strong learned baseline. Raw ICI and Raw Mel are feature floors (logistic regression on raw features with no learned encoder) — useful as sanity references, but not "strong baselines" in the sense the feedback asks for. (Logical: a linear model on raw features is by construction a lower bound, not a competitive method.)
• Consequence: the paper currently compares DCCE against essentially one external model. The plan below adds 1–3 strong, literature-justified baselines.

2. The principle we use to decide what counts as a "strong" baseline

(This section is a logical/methodological argument, not a citation — stated explicitly so the criteria are transparent.)

A baseline is "strong" for our paper if it satisfies all three:

1. It is a learned representation established in the literature for audio / bioacoustic tasks (not a trivial feature transform).
2. It has precedent for our setting — ideally used in the literature for the same task (individual identification) or in the same comparison context (e.g., it was a baseline in the WhAM paper we build on).
3. It is runnable on our exact data and evaluation protocol (frozen embeddings + linear probe on the identical 80/20 split), so the numbers are directly comparable.

This filter is what separates "strong baselines worth doing" from "weak baselines not worth doing."

3. The baselines we propose to add (each with its basis in the literature)

3.1 Keep WhAM as the primary integrated baseline

• WhAM is the current state-of-the-art learned representation for this exact dataset, and our work explicitly builds on it (Paradise et al., WhAM, NeurIPS 2025, arXiv:2512.02206).
• Logical point: being inspired by WhAM does not disqualify it as a baseline — it makes it the most necessary one. If we build on WhAM and never compare to it, the obvious question "is DCCE actually better than just using WhAM embeddings?" goes unanswered. (Methodological rationale.)

3.2 AVES — bioacoustic self-supervised encoder (primary new baseline)

• AVES is a self-supervised animal-vocalization encoder (HuBERT-based) designed for bioacoustic transfer (Hagiwara, AVES, ICASSP 2023, arXiv:2210.14493).
• Why it is strong / has precedent: WhAM's own evaluation uses AVES as a baseline — so it is the field's established category-peer for "a learned representation of animal sound." (Paradise et al. 2025; corroborated in our literature notes.)
• What it tests: whether DCCE's purpose-built dual-channel representation beats a general-purpose self-supervised animal-audio encoder.

3.3 BirdNET / Perch — bioacoustic transfer embeddings (best fit for our headline task)

• BirdNET is a widely deployed bird-sound classifier whose embeddings are routinely reused for transfer (Kahl et al., BirdNET, Ecological Informatics 61:101236, 2021); Perch embeddings give state-of-the-art bioacoustic transfer (Ghani et al., Global Birdsong Embeddings, Sci. Rep. 13:22876, 2023, arXiv:2307.06292).
• Why it is strong / has precedent for our exact task: Tolkova et al. (2025, bioRxiv 10.1101/2025.07.11.664307) used BirdNET embeddings + a shallow classifier to identify individual North Atlantic right whales (11 individuals, small dataset) — almost exactly our task and data regime. This is the strongest possible literature precedent for "transfer embeddings as a baseline for cetacean individual ID."
• Additional support: in WhAM's own downstream evaluation, BirdNET embeddings were a strong baseline (e.g., ~92% on social unit), sometimes rivaling WhAM. (Paradise et al. 2025 — their reported table.)
• Note for the professor: BirdNET/Perch run on TensorFlow, so this baseline requires extra setup; we flag it as high-value-but-higher-effort.

3.4 wav2vec 2.0 / HuBERT — generic audio SSL (off-the-shelf control)

• Canonical self-supervised speech encoders (Baevski et al., wav2vec 2.0, NeurIPS 2020, arXiv:2006.11477; Hsu et al., HuBERT, IEEE/ACM TASLP 2021, arXiv:2106.07447).
• Why include it: Sarkar & Magimai-Doss (Interspeech 2023, arXiv:2305.14035) showed that human-speech SSL features carry individual-caller information for animals without fine-tuning — so a generic SSL encoder is a meaningful "did you even need a domain model?" control.
• Practical note: these load directly via torchaudio, so they are the lowest-effort of the new baselines.

3.5 Why we do not add a stronger rhythm-channel baseline

• A logistic regression on raw ICIs already reaches macro-F1 0.931 on coda type. (Our own Phase 1 result.)
• Logical conclusion: coda type is by definition determined by the ICI timing pattern (Gero et al., R. Soc. Open Sci. 2016), so the rhythm channel is already near its ceiling and a more complex rhythm baseline would add little. We therefore concentrate the "strong baselines" on the spectral / learned-audio side, and say so explicitly.

4. Where we conducted the research (sources)

We ran a structured literature scan across four adjacent bodies of work, fetching from arXiv, bioRxiv, Nature / Nature Communications, Scientific Reports, Royal Society Open Science / Proc. R. Soc. B, and the ICASSP / Interspeech proceedings. The four areas:

• Acoustic individual identification across animals — to establish that the task is real and that learned-embedding baselines exist for it. Key sources: Carlson et al. review (Phil. Trans. R. Soc. B 2020, DOI 10.1098/rstb.2019.0479); Tolkova et al. 2025 (right whales); Nolasco et al. 2024 (hierarchical contrastive learning for individual ID, arXiv:2409.08673); Bermant et al. 2019 (first deep learning on sperm whale codas, Sci. Rep. 9:12588).
• Bioacoustic foundation models — to identify the strong domain encoders: AVES, BirdNET, Perch (cited above); BEANS benchmark (Hagiwara et al., ICASSP 2023, arXiv:2210.12300).
• General self-supervised audio learning — for the generic baselines and the evaluation convention: wav2vec 2.0, HuBERT (cited above).
• Probing methodology — to justify the frozen-embedding + linear-probe protocol we use to compare all models on equal footing: Pasad et al. (layer-wise SSL probing, ASRU 2021, arXiv:2107.04734) and Castellon et al. (JukeMIR, ISMIR 2021, arXiv:2107.05677), the precedent for probing a generative model's internal layers (exactly what we do to WhAM).

(Full per-paper notes with verified DOIs/IDs are in reports/dcce_literature_foundations.md. A few preprints — Tolkova 2025, Nolasco 2024 — are flagged there as not-yet-peer-reviewed.)

5. The fairness consideration that makes the comparison valid

(Primarily a methodological argument, with supporting literature.)

• The asymmetry: DCCE-full learns its representation with individual-ID label supervision (the L_id auxiliary loss), whereas WhAM, AVES, and the SSL encoders never see ID labels. Comparing them directly on individual ID is therefore not apples-to-apples. (Logical.)
• Why this matters in practice: our own ablations show the cross-channel novelty moves individual-ID macro-F1 by only +0.009 (late_fusion 0.825 → full 0.834), while the supervised single-channel spectral_only already reaches 0.787. So most of the headline ID number comes from supervision, not the architectural novelty. (Our own Phase 3 ablation results.)
• Established concern: disentangling identity from confounding/training signals is treated as a first-class problem in the speaker-recognition SSL literature (e.g., contrastive speaker learning with channel-invariance: Huh et al. 2020, arXiv:2007.12085; Zhang et al. 2021, ICASSP). We import the same rigor.
• The fix: report results in tiers by supervision — a "label-free representations" tier (WhAM, AVES, SSL, and a DCCE variant trained without L_id) versus an "ID-supervised" tier (DCCE-full and supervised controls). The fair test of our contribution lives in the label-free tier.

6. How we aim to expand (the plan)

1. Run the strong baselines on our exact split. Extract frozen embeddings (AVES, wav2vec2/HuBERT; BirdNET/Perch if we add TensorFlow) and evaluate with the identical linear-probe protocol already used for WhAM, so every number is directly comparable. (Standard SSL evaluation: SimCLR linear-probe convention, Chen et al. ICML 2020; layer-wise probing, Pasad et al. 2021.)
2. Add the fairness tier. Train a DCCE variant without L_id so our architecture also appears in the label-free comparison group.
3. Add error bars. Report mean ± standard deviation of macro-F1 over repeated stratified splits, because the individual-ID test set has small per-class counts and single-split point estimates are unstable. (Logical: small-sample variance; standard practice.)
4. Present results in a tiered table (feature floors → strong single-channel baselines → SOTA integrated baseline → DCCE), making the rationale explicit.
5. Future expansion (optional): evaluate retrieval / few-shot generalization under distribution shift, following the bioacoustic retrieval benchmark BIRB (Hamer et al. 2023, arXiv:2312.07439) — relevant because our data carries a recording-year confound.

7. Summary: baseline → basis → what it tests

One-paragraph version (for speaking to it)

Our original baselines were weak: only WhAM was a genuine learned comparison, the others were raw-feature floors, and we had blurred baselines with ablations. We are fixing this by adding strong baselines that the bioacoustics literature already endorses for our task: AVES, the self-supervised animal-audio encoder that WhAM itself benchmarked against; and BirdNET/Perch embeddings, which Tolkova et al. (2025) used to identify individual right whales on a dataset much like ours. We keep WhAM as the state-of-the-art integrated baseline, but compare it fairly by also training a version of our model without individual-ID supervision, since WhAM never saw those labels. All baselines are evaluated with the same frozen-embedding linear-probe protocol on the same split, so the comparison is apples-to-apples.

Shuo's Suggestion

• Overleaf suggestion - put your paper in Overleaf and share that with the three of us so that they can all edit
• This way Doctor Shuo's can comment more and either edit and do it there properly
• You need to download the DCASE template to fit
• For appendix -> typically it is okay having several pages
• Emma has been working on shortening the literature review and cleaning that up.