Talk2NatureSperm whale · cetacean bioacoustics / Generative audio and representation learning
WhAM: acoustic style transfer is not whale translation
arXiv v1 posted 2025-12-01; official project repository identifies the work as NeurIPS 2025 2025 · Note reviewed 2026-10-13
WhAM fine-tunes the music-pretrained VampNet masked acoustic token model in two stages: domain adaptation on animal-audio collections, then sperm-whale-specific training on DSWP and CETI. The paper describes 2,507 DSWP and 7,653 CETI annotated codas for this work. It evaluates generated audio with Fréchet Audio Distance and a five-expert listening study, and tests learned embeddings on coda detection, rhythm, social-unit and vowel-type classification. The paper explicitly defines “translation” as converting prompts, including noise or other animal sounds, into sperm-whale coda-like acoustic style; it says semantic translation remains a separate, more ambitious goal. On the reported classification table, WhAM exceeds simple random/majority baselines but the reported specialist or general bioacoustic embeddings perform better across these tasks. For example, social-unit accuracy is 70.5% for WhAM versus 93.2% for BirdNET. The paper describes social-unit classification within data from a single vocal clan as closer to speaker identification than dialect classification.
What we reviewed · Full main-text review and targeted methods/artifact metadata review; no independent reproduction
Read the full arXiv v1 main text from abstract through limitations, and reviewed targeted Appendix E sections on training data, compute, downstream classification and expert perceptual evaluation. Checked the official repository README/license and pinned source revision, the DSWP dataset card and its current repository revision, and the Zenodo model-weight record. No code, audio, datasets, model weights, paper figures or supplementary files were downloaded or run; Appendix D and artifact contents were not fully audited.
No independent cetacean-bioacoustics specialist review, author verification or Talk2Nature reproduction has occurred.
What this does — and doesn’t — establish
The downstream protocol describes an 80/20 label-stratified train/test split and selects the best classifier checkpoint based on test-set performance. The reviewed description does not establish a separate validation set or grouping by whale, recording, session or source, so the reported test metrics should not be treated as an untouched independent estimate. The public DSWP card currently lists 1,501 isolated snippets (~45 minutes; 585 MB), fewer than the paper’s 2,507 DSWP training codas; the exact relationship is unresolved. The card says the clips lack behavioral context and per-file recording-system metadata and all come from one clan and region. The paper identifies codec limitations, unnatural synthetic click onset/decay/background, echolocation-like artifacts and possible echolocation contamination in data prepared as communication codas. Its expert study involved five specialists with only moderate inter-expert agreement. The acoustic outputs and classification labels do not demonstrate receiver comprehension, behavioral function or meaning. The paper is CC BY-NC-ND 4.0, code MIT, DSWP card CC BY 4.0, while the weight record displays no license; these are distinct component rights. No independent reproduction or specialist review occurred.
What we can build on
Use WhAM to study a specific generative-modeling and representation-learning approach, not as a conversation engine or a reason to play generated animal-like sounds to animals. Its test-set model selection and unspecified entity grouping are concrete examples for Talk2Nature’s benchmark rules: group correlated samples before splitting, freeze an untouched test set, select checkpoints only on validation data, compare simple and domain-specific baselines, and keep acoustic similarity separate from biological function or meaning.
See the related tools and reading paths →
The next reading task
If this becomes relevant to a future benchmark, first obtain qualified cetacean-bioacoustics review, resolve exact DSWP/CETI release provenance and rights, verify a grouped split by relevant animal/recording/session, and freeze validation and test protocols before any model evaluation. Resolve the model-weight license and dependency/data terms before download or execution. Do not use it in the public companion or initiate playback. AutoResearch remains NOT READY.
Go to the sources
- WhAM primary paper, full-text review ↗
Paradise et al., arXiv:2512.02206v1 (2025), NeurIPS 2025. Full main text and selected methods appendices reviewed. The paper explicitly defines “translation” as acoustic style transfer; semantic translation remains a distinct goal.
- Project CETI WhAM code at pinned GitHub revision ↗
Official WhAM repository resolved to commit 00a8b787c040db23cd51ac4417481a09ac354985 on 2026-10-13. GitHub identifies the repository license as MIT.
- Public DSWP dataset card used by the WhAM project ↗
The current Hugging Face card lists 1,501 Dominica sperm-whale codas (~45 minutes; 585 MB) and declares CC BY 4.0. The WhAM paper reports 2,507 DSWP codas in its training description; the relationship between these counts is unresolved.
- WhAM model weights, Zenodo version 1 ↗
Zenodo DOI 10.5281/zenodo.17633708, published 2025-12-02. The record lists four weight files totaling about 3.1 GB; its rendered Rights/License field is blank.
Original editorial note by Talk2Nature contributors. AI-assisted, source-grounded synthesis; not a peer review or an independent replication. Linked works retain their own rights.