Reading list, organized by constellation.

Marzluff's Seattle research program is the empirical bedrock for almost every modern claim about crow individuality, memory, and social cognition. This book translates the lab work into a public-facing account without losing rigor — and it remains the entry point most professional ornithologists recommend to first-time corvid readers. CrowLingo's territorial caw cluster narrative cites it directly.

Heinrich's slow naturalist work on common ravens shows what a decade of patient observation reveals that no instrumented study yet can — including individual personality, cross-generational learning, and play behavior. Crows are a separate species, but the cognitive architecture rhymes. Any honest ALP project starts with this book as a check on its own assumptions about what's actually happening behind the call.

The classical-bioacoustics result that motivates the whole vocal-atlas approach: even a single caw type — when measured with hand-crafted spectral features — separates caller sex, individual identity, and broad behavioral context. ALP doesn't replace this finding; it extends it to a continuous embedding space where the same separations emerge geometrically. Mates et al. is what makes the territorial cluster's claim of 'identity packed in milliseconds' an established fact, not speculation.

The 2024 update of the Birds of the World monograph is the closest thing the field has to a canonical reference text on American crow vocalizations. Every cluster name on CrowLingo's atlas — territorial, mobbing, assembly, rattle, begging, exceptional — descends from terminology this monograph standardizes. Cornell's audio library powers the embedded examples; their license terms are why we built the v1 corpus from Wikimedia mirrors instead.

The April 2026 preprint that defines the new methodological standard: instead of pointing microphones at colonies, wearable bioacoustic loggers on individual carrion crows captured 127,000+ vocalizations across behavioral contexts. Demartsev et al. then mapped the full repertoire with Voxaboxen and UMAP, recovering both discrete and graded structure that fifty years of hand-labeling missed. CrowLingo's atlas is the public-facing American crow analog of this methodology.

Dan Stowell's review is the single most-cited overview of where bioacoustics was when deep learning arrived. It catalogues which problems are solved (species ID), which are progressing (individual ID, behavioral classification), and which are still open (graded variation, semantic interpretation). If you only read one paper before judging whether ALP is overhyped or under-hyped, this is the one. CrowLingo treats it as the operating manual for what's actually defensible.

The textbook that defines what 'communication' even means in a non-human context: signal, receiver, channel, code, evolutionary stable strategy. Before you can claim a crow call has meaning, you need this framework to specify what you mean by meaning. Bradbury & Vehrencamp is unfashionable to cite — it's a textbook, not a paper — but every careful claim on CrowLingo is implicitly Bradbury-and-Vehrencamp-compliant.

BirdNET is the open-source bird-detection model that ended up running on every phone, every smart speaker, every long-term acoustic monitoring deployment. It's not the most sophisticated model anymore — Perch 2.0 outperforms it — but its training corpus and evaluation methodology shaped what 'good' means in the field. CrowLingo's atlas embeddings are downstream of design choices Kahl et al. made in this paper.

The Earth Species Project's ICLR 2025 paper is the moment audio foundation models stopped being a vision-borrowed metaphor and became a bioacoustics-native tool. NatureLM-audio combines a BEATs audio encoder with a Llama-3.1 language backbone, trained on a curated bioacoustic corpus, and answers natural-language questions about audio clips zero-shot. The atlas page's NatureLM-audio attribution is not branding — it's the model whose embeddings would back a production version of our similarity-search feature.

BEANS standardized how to compare audio foundation models on bioacoustic tasks — species classification, call detection, individual ID — across a curated set of 12 sub-tasks spanning birds, mammals, insects, and amphibians. Before BEANS, every paper picked its own evaluation; after BEANS, the SOTA leaderboard is legible. BEANS-Zero is the zero-shot extension that defines whether a model has actually learned bioacoustics-general features or just memorized its training distribution.

Earth Species Project's launch post for NatureLM-audio is the cleanest public framing of the ALP research agenda: catalog, classify, characterize, eventually decode — explicitly in that order, explicitly with humility about the last step. It also names the failure modes — anthropomorphism, over-reach, ecological harm — that any ALP project has to design against. CrowLingo's own ethics floor descends from this framing.

CETI is the most ambitious public-facing ALP project: a five-year, multi-million-dollar effort to decode sperm whale codas using language models, photogrammetry, and behavioral data. The Andreas et al. roadmap paper is unusual for being honest about the project's epistemological risks: that 'translation' may be the wrong frame, that meaning may not be lexical, that we may discover absence of structure where popular coverage assumed presence. CrowLingo's editorial stance on 'don't promise translation' is downstream of reading this paper carefully.