Publications

Argues that acoustic indices have a flawed theoretical basis, weak empirical support, and poor interpretability as biodiversity proxies, and makes the case that acoustic monitoring should instead rely on species-level data from manual annotation, single-species detectors, and multi-species classifiers.

Applies passive acoustic monitoring across 25,000 km² of Sierra Nevada conifer forest to study the occupancy dynamics of Clark's nutcracker, showing how conifer composition, elevation, and recent severe fire shape where this nomadic keystone seed disperser occurs.

Based on a structured solicitation of 30 international PAM practitioners, identifies and ranks the main software pain points in acoustic workflows, led by accurate AI-based species identification in noisy soundscapes, and proposes practical strategies to address fragmented tools, validation bottlenecks, and data-sharing challenges.

BirdCLEF+ 2026 moves to the Brazilian Pantanal, the world's largest continental wetland, and asks participants to identify birds, amphibians, insects, mammals, and reptiles in continuous soundscapes under a CPU-only inference limit. More than 4,000 teams took part, and the best systems combined distilled and ensembled bioacoustic foundation models, pseudo-labeling, and ecological priors, reaching a ROC-AUC of 0.965.

Shows how passive acoustic surveys analyzed in a Bayesian dynamic occupancy framework can robustly estimate occupancy dynamics and responses to disturbance in the California spotted owl, which is threatened by increasingly large, severe megafires.

Develops a custom BirdNET classifier for the critically rare orange-bellied parrot in Tasmania and shows that acoustic surveys achieved a detection probability more than eleven times higher than field surveys by experienced observers, providing a template for monitoring rare, mobile species.

Uses BirdNET and recordings from across the Sierra Nevada to show that avian vocal output drops after goshawk calls, and retrains BirdNET to separate mountain chickadee call types, revealing habitat-dependent antipredator behavior that could only be detected at a regional scale.

Demonstrates a passive acoustic monitoring system that records and classifies audio at remote sites and reports BirdNET detections of at-risk birds to land managers in near real time at Pōhakuloa Training Area, Hawaiʻi, and identifies wind and non-target vocalizations as the main sources of detection errors.

Using a proposed acoustic monitoring program for the threatened Mexican spotted owl as an example, shows that sample size recommendations from occupancy-based power analyses are highly sensitive to how trends are simulated and how power is checked, and translates these choices into cost-optimized monitoring designs.

Offers suggestions for how terrestrial passive acoustic monitoring programs can proactively plan for extension (growth over time), expansion (growth over space or in the species surveyed), and integration (growth in the number of teams involved, leading to synthesis of data across programs) to maximize research and conservation impact.

Models the vocal activity rates of 94 bird species at 553 acoustic monitoring stations in the Sierra Nevada using 129 geospatial layers, showing that emerging remote sensing data such as lidar-derived forest structure and hyperspectral foliar traits significantly improve predictions of species distributions.

Combines a trait-based selection of functionally representative, acoustically detectable bird species with passive acoustic monitoring in north-central Namibia, and uses occupancy models to show how land management, soil, and vegetation shape their distributions across a dryland land use mosaic.

Uses four years of landscape-scale acoustic monitoring and dynamic occupancy models to show that the Hermit Warbler declined by 8.3%, shifting away from warmer, lower-elevation sites and failing to recolonize severely burned areas, which indicates that both fire and rising temperatures threaten this old-forest indicator.

Uses passive acoustic monitoring during the 2023 Canadian wildfire smoke events to show that heavier smoke reduced the vocal activity of breeding grassland birds in New York State, suggesting that smoke is an additional stressor for already declining species.

Evaluates how well surrogate species capture spatial patterns of biodiversity across an ecosystem, finding that surrogate effectiveness varies geographically and emphasizing the need for region-specific validation when using umbrella species.

Describes the "Bioacoustics on Tiny Hardware" task of the first BioDCASE challenge, which encouraged the development of bioacoustic detection and classification methods that run on resource-constrained embedded devices.

Demonstrates that large-scale passive acoustic monitoring networks can provide spatially continuous information on bird communities that is useful for planning fire and forest restoration treatments across bioregional scales.

Combines data from a large network of autonomous recorders and citizen-science devices to show that bird vocal activity declines sharply only in areas experiencing near-total obscuration during a solar eclipse, revealing large-scale behavioral responses to transient darkening.

Describes the deployment of a continuous real-time acoustic monitoring system for endangered Hawaiian forest birds and shows how streaming detections can inform management and research in remote, logistically challenging environments.

Shows that native predator species differ in their responses to severe wildfire and invasive prey according to life-history traits, highlighting that management actions may differentially affect predator guild members.

Demonstrates that passive acoustic monitoring is an effective approach for surveying wildlife species and communities in the southwestern United States, laying the groundwork for a regional program that tracks species and community trends in response to multiple stressors.

Shows that owl assemblages in Sierra Nevada forests are maintained under frequent, heterogeneous fire regimes, suggesting that restoring pyrodiversity can benefit multiple predator species simultaneously.

Analyzes the impact of label noise in training data on BirdNET’s performance, showing that mislabeled examples can substantially degrade accuracy and providing recommendations for curation and robust training strategies.

BirdCLEF+ 2025 expands the task beyond birds, introducing multi-taxonomic acoustic identification in Colombia’s Middle Magdalena Valley. The challenge evaluates systems on birds, mammals, amphibians, and insects, reflecting the increasing role of generalized ecoacoustic monitoring. Early results suggest that cross-taxa models benefit from shared feature representations but struggle with uneven data availability.

Compares species-specific and universal confidence thresholds for BirdNET detections, showing that species-specific thresholds generally yield better classification performance while universal thresholds may be acceptable where data are limited, and offers practical recommendations for threshold selection in ecological studies.

Integrates passive acoustics, camera traps and remote sensing into a unified analytical framework to quantify biodiversity change following large disturbances, demonstrating a transferable approach for multi-taxa monitoring at landscape scales.

Presents BirdSet, a curated collection of multiple labeled bird sound datasets with standardized splits and multi-task evaluation protocols to benchmark classification models across diverse acoustic and biogeographic conditions.

Uses automated detection and embedding-based analyses to identify a previously unrecognized red crossbill call type in northeastern North America, highlighting the role of machine learning in uncovering cryptic vocal diversity.

Clarifies how to interpret and use detector scores such as those from BirdNET, emphasizing calibration, species- and context-specific performance, and the need to propagate uncertainty when drawing ecological inferences.

Systematically evaluates combinations of spectrogram-based audio augmentation techniques for birdsong classification and shows that carefully designed augmentation pipelines can substantially improve generalization on imbalanced and noisy training data.

BirdCLEF 2024 focuses on the Western Ghats, a global biodiversity hotspot with many understudied and data-poor species. The paper presents newly collected soundscape data and evaluates machine learning systems on sparse training regimes. Results show that transfer learning and multi-species modeling are important for effective identification in low-resource settings.

Deploys a real-time acoustic monitoring system for an invasive bird species and shows that near-real-time detections can be used operationally to guide management actions and reduce response latency.

Shows that integrating passive acoustic monitoring into spotted owl survey protocols can increase detection efficiency and reduce field effort, with implications for designing restoration treatments in fire-affected forests.

Repurposes BirdNET to detect and classify wolf and coyote vocalizations in extensive passive acoustic datasets, illustrating cross-taxon application and documenting patterns that may indicate interspecific interactions.

Curates annotated soundscapes from shade-coffee farms in Colombia and Costa Rica, supporting work on how agricultural management influences acoustic biodiversity and ecosystem services.

Provides annotated soundscape recordings from the southern Sierra Nevada, facilitating evaluation of acoustic monitoring for forest birds in fire-affected montane landscapes.

Introduces BIRB, a benchmark for audio retrieval of bird vocalizations under distribution shift, and shows that current representation-learning approaches struggle with cross-dataset and cross-task generalization in realistic bioacoustic settings.

Reviews published applications of BirdNET in ornithological research, assessing its performance across habitats, species, and study designs, and highlights common pitfalls such as species coverage gaps, context-dependent accuracy, and threshold selection, along with best practices and future directions for monitoring and conservation.

Uses large-scale passive acoustic monitoring and calling-rate models to estimate population sizes of California spotted owls and barred owls, demonstrating how bioacoustics can support landscape-scale population assessments of elusive raptors.

Demonstrates that BirdNET embeddings capture ecologically meaningful variation in avian communities, showing that embedding-based analyses can reveal gradients in habitat, disturbance and community composition beyond simple species lists.

Trains global birdsong embedding models on large, diverse datasets and demonstrates that these embeddings substantially outperform baseline features for transfer-learning tasks across species, regions and recording conditions.

BirdCLEF 2023 introduces a new large-scale dataset from Eastern Africa, emphasizing biodiversity-rich tropical ecosystems. The challenge encourages models that generalize across sites, species, and recording conditions. Findings highlight persistent gaps in performance for rare and acoustically similar species.

Applies a BirdNET-based detector to primate vocalizations in passive acoustic data and shows that detector outputs can be integrated into occupancy models to estimate site use of an endangered primate species.

Combines passive acoustic monitoring with BirdNET-based detection to characterize fine-scale spatial and temporal patterns in calling activity of two frog species, illustrating the transferability of bird-focused models to other taxa.

Uses occupancy modeling and fire-history data to show that mountain quail in fire-adapted forests can benefit from frequent, heterogeneous fire regimes, highlighting how restoration of historical fire patterns may aid some disturbance-adapted species.

Provides annotated soundscape recordings from the Island of Hawaiʻi, focusing on native and non-native species, and supporting research on tropical island bioacoustics and invasive species impacts.

Provides a fully annotated soundscape dataset from Northeastern United States sites, including detailed species-level labels suitable for benchmarking bird sound detection, classification and soundscape analysis methods.

Offers a labeled soundscape dataset from the southwestern Amazon Basin, capturing high-diversity tropical bird communities and enabling assessments of automated acoustic monitoring methods in complex soundscapes.

Releases annotated soundscape recordings from Western United States sites with species-level labels, enabling evaluation of automated detectors and ecological analyses in fire-prone montane forest ecosystems.

BirdCLEF 2022 shifts attention toward endangered species detection, introducing datasets from conservation-critical regions such as Hawai‘i. The paper details the task setup and evaluation criteria designed to support real-world conservation monitoring. System results underscore the difficulty of rare-species classification in noisy habitats.

Evaluates the BirdNET mobile app as a citizen-science tool and shows that machine-learning-based bird sound identification can substantially expand spatial and taxonomic coverage of avian observations, especially in under-sampled regions.

Introduces BirdNET, a deep convolutional network trained on large-scale annotated bird sound data, and shows that it can reliably detect and identify hundreds of species, enabling scalable passive acoustic monitoring for biodiversity studies.

BirdCLEF 2021 refines soundscape-based bird call detection with new data sources and improved annotations. It evaluates model performance on varied ecological contexts, emphasizing robustness in passive acoustic monitoring applications. Results show strong gains from ensembles and advanced data augmentation strategies.

Uses deep audio embeddings derived from BirdNET to segment and cluster song elements within continuous birdsong recordings, showing that generic bioacoustic embeddings can support fine-grained, unsupervised structure discovery in vocal sequences.

Quantifies how spatial coverage, recording duration and community composition influence species richness estimates from passive acoustic surveys, providing guidance on survey design to reduce underestimation and improve comparability across studies.

BirdCLEF 2020 focuses on automated identification of birds in highly complex acoustic soundscapes, featuring overlapping species and environmental noise. The authors outline the challenge design, curated datasets, and training resources. System comparisons highlight continued improvements in deep-learning-based approaches.

Develops BirdNET as a comprehensive deep learning framework for large-scale avian acoustic event recognition and demonstrates its applicability for automated biodiversity monitoring and long-term avian activity assessments across diverse recording conditions.

BirdCLEF 2019 expands the challenge toward larger-scale soundscape recognition with increased species diversity and geographic coverage. The paper describes the dataset, annotation protocols, and evaluation metrics used to benchmark automated systems. Results emphasize the growing challenge of species-rich and acoustically complex habitats.

BirdCLEF 2018 compares single-species audio classification with full soundscape-based bird identification. The paper introduces the dataset, evaluation setup, and system performance across participating teams. The work highlights the difficulty of polyphonic and noisy environmental recordings in real-world biodiversity monitoring.

Presents a convolutional neural network approach for large-scale bird sound classification in the BirdCLEF 2017 challenge, using spectrogram-based inputs and extensive data augmentation to achieve competitive mean average precision on highly imbalanced training data.