Search

Convolutional neural networks (CNNs) are effective tools for acoustic classification tasks such as species identification. Large datasets of labelled recordings are required to develop CNN classifiers which can be difficult to obtain, particularly if species are rare or vocalise infrequently. Additionally, data often requires manual labelling which can be time consuming requiring expert analysis. Artificially generating data using augmentation can address these challenges, however the impact of data augmentation on CNN performance is poorly understood and often omitted in bioacoustic studies. Here, we empirically test the impact of CNN architecture and 20 data augmentation methods on classifier performance. We use acoustic identification of 18 small mammal species as a case study of a species group that can be effectively surveyed by acoustic monitoring, but recordings for training data are scarce and difficult to collect. Networks that achieved the highest accuracy across all sample sizes was a 10-layer CNN (96.43 %) and a pre-trained ResNet50 model (96.37 %). Overall, all augmentation effects improved ResNet50 model performance and 17 effects improved Conv10 performance, increasing relative change in accuracy (RCA) by 0.021–0.641. Three augmentation effects negatively impacted Conv10 RCA by −0.042 to −0.182. We also show that adding augmented data when the number of original samples is low has the greatest positive impact on accuracy and this effect was larger with ResNet50 models. Our work demonstrates that using data augmentation where few original samples are available can considerably improve model performance and highlights the potential of augmentation in developing acoustic classifiers for species where data are limited or difficult to obtain.

The collection and annotation of bioacoustic data present several challenges to researchers. Bioacoustic monitoring of rare (sparse) or cryptic species generally encounter two main issues. The cost of collecting and processing field data and a lack of labelled datasets for the target species. The detection of invasive species incursions and probability of absence testing is especially challenging due to these species having population densities at or close to zero. We present a methodology specifically designed to aid in the analysis of rare acoustic events within long-term field recordings. This approach combines a wavelet-based segmentation method that automatically extracts transient features from within-field recordings. A few-shot active learning recommender system in a human-in-the-loop process prioritises the annotation of low-certainty samples. This process combines the accuracy of human classification and the speed of computational tools to greatly reduce the presence of non-target features in field recordings. We evaluate this approach using an invasive species identification case study. This methodology achieves a test accuracy of 98.4% as well as 81.2% test accuracy using 2-shot, 2-way prototypical learning without fine-tuning, demonstrating high performance at varying data availability contexts. Active learning using low-certainty samples achieves >90% test accuracy using only 20 training samples compared to 80 samples without active learning. This approach allows users to train custom audio classification models for any application with rare features. The model can be easily exported for use in the field making real-time bioacoustic monitoring of less-vocal species a possibility. All code and data are available at https://github.com/Listening-Lab/Annotator.

Convolutional neural networks (CNNs) are effective tools for acoustic classification tasks such as species identification. Large datasets of labelled recordings are required to develop CNN classifiers which can be difficult to obtain, particularly if species are rare or vocalise infrequently. Additionally, data often requires manual labelling which can be time consuming requiring expert analysis. Artificially generating data using augmentation can address these challenges, however the impact of data augmentation on CNN performance is poorly understood and often omitted in bioacoustic studies. Here, we empirically test the impact of CNN architecture and 20 data augmentation methods on classifier performance. We use acoustic identification of 18 small mammal species as a case study of a species group that can be effectively surveyed by acoustic monitoring, but recordings for training data are scarce and difficult to collect. Networks that achieved the highest accuracy across all sample sizes was a 10-layer CNN (96.43 %) and a pre-trained ResNet50 model (96.37 %). Overall, all augmentation effects improved ResNet50 model performance and 17 effects improved Conv10 performance, increasing relative change in accuracy (RCA) by 0.021–0.641. Three augmentation effects negatively impacted Conv10 RCA by −0.042 to −0.182. We also show that adding augmented data when the number of original samples is low has the greatest positive impact on accuracy and this effect was larger with ResNet50 models. Our work demonstrates that using data augmentation where few original samples are available can considerably improve model performance and highlights the potential of augmentation in developing acoustic classifiers for species where data are limited or difficult to obtain. • A CNN-based small mammal acoustic classifier was developed to investigate the impact of data augmentation on performance. • Augmenting data can improve classifier accuracy by up to 8.4 % when there are as few as 100 samples per species. • Using augmentation to balance classes with uneven sample sizes reduces classification errors for underrepresented species. • We show that data augmentation can reducce the need for large, labelled datasets for deep learning acoustic classifiers. [ABSTRACT FROM AUTHOR]

The Southern Resident Killer Whale (SRKW) is an endangered population of whales found in the northeast Pacific. They have a vocal dialect unique from other killer whales, having a repertoire of distinct stereotyped calls. A framework for distinguishing SRKW call types using the frequency traces of the amplitude ridges from their spectrograms (termed frequency ridges) is proposed. The first step is the extraction of these ridges of SRKW calls using an Sequential Monte Carlo approach. Next, they are converted into functional data using B-spline functions. They are analyzed with a functional principal component (FPC) analysis to characterise the intrinsic variability of frequency ridges within a call type. The FPCs are able to capture the general patterns in the frequency ridges of the different SRKW call types. The FPCs are also used as the basis for call classification. Using a cross-validation procedure to assess the robustness of the classification, this framework proves to be successful for classification with some call types having an F1-score ≥ 80 % , but other calls less well discriminated. On balance, this approach showed reasonable performance given the small sample size available, and provides a useful contribution towards the development of a universal tool for call classification. • The Southern resident killer whale is endangered. • Clues on their call types resides in their frequency ridge shapes in the spectrogram. • Functional data analysis can be used to classify frequency ridges into call types. • Functional data analysis offers a novel approach for acoustically distinguishing pods. [ABSTRACT FROM AUTHOR]

The collection and annotation of bioacoustic data present several challenges to researchers. Bioacoustic monitoring of rare (sparse) or cryptic species generally encounter two main issues. The cost of collecting and processing field data and a lack of labelled datasets for the target species. The detection of invasive species incursions and probability of absence testing is especially challenging due to these species having population densities at or close to zero. We present a methodology specifically designed to aid in the analysis of rare acoustic events within long-term field recordings. This approach combines a wavelet-based segmentation method that automatically extracts transient features from within-field recordings. A few-shot active learning recommender system in a human-in-the-loop process prioritises the annotation of low-certainty samples. This process combines the accuracy of human classification and the speed of computational tools to greatly reduce the presence of non-target features in field recordings. We evaluate this approach using an invasive species identification case study. This methodology achieves a test accuracy of 98.4% as well as 81.2% test accuracy using 2-shot, 2-way prototypical learning without fine-tuning, demonstrating high performance at varying data availability contexts. Active learning using low-certainty samples achieves > 90% test accuracy using only 20 training samples compared to 80 samples without active learning. This approach allows users to train custom audio classification models for any application with rare features. The model can be easily exported for use in the field making real-time bioacoustic monitoring of less-vocal species a possibility. All code and data are available at https://github.com/Listening-Lab/Annotator. • Developed an open-source audio analysis methodology designed for bioacoustic monitoring applications with rare features. • Present a methodology that integrates few-shot learning and active learning for efficient analysis of sparse features. • Evaluate methodology using invasive species detection case study. [ABSTRACT FROM AUTHOR]

Automatic acoustic monitoring of bird populations and their diversity is in demand for conservation planning. This requirement and recent advances in deep learning have inspired sophisticated species recognizers. However, there are still open challenges in creating reliable monitoring systems of natural habitats. One of many open questions is whether predominantly used audio features like mel-filterbanks are appropriate for such analysis since their design follows human's perception of the sound, making them susceptible to discarding fine details from other animals' vocalization. Although research shows that different audio features work better for particular tasks and datasets, it is hard to attribute all advantages to input features since the experimental setups vary. A general solution is to design a learnable audio frontend to extract task-relevant features from raw waveform since it contains all the information in other audio features. The current paper thoroughly analyzes the role of such frontends in bird species recognition, which helped to evaluate the adequacy of traditional time-frequency representations (static frontends) in capturing the relevant information from bird vocalization. In particular, this work shows that the main performance gain in learnable audio frontends comes from the normalization and compression operations rather than the data-driven frequency selectivity and functional form of filters. We observed no significant discrepancy between the frequency bands of the learned and static frontends for bird vocalization. Although the performance of learnable frontends was much higher, we will show that adequate normalization and compression enhance the accuracy of traditional frontends by more than 16% to achieve comparable results for bird species recognition. Ablation studies of the frontends under different configurations and detailed analysis of noise robustness provide evidence for the conclusions, validate the use of mel-filterbanks and similar features in prior works, and provide guidelines for designing future species recognizers. The code is available at https://github.com/houtan-ghaffari/bird-frontends. • Learnable audio frontends impact species classifiers significantly. • Learnable filters are log-linearly spaced for bird vocalization like mel-filterbank. • PCEN and normalization make mel-spectrogram and learnable filters comparable. • Compression and normalization make the models robust against environmental noise. • Filter parametric form is significant without normalization and smoothing modules. [ABSTRACT FROM AUTHOR]

Deep neural networks (DNN) are a popular tool to process environmental sounds and identify sound-producing animals, but it can be difficult to understand the decision-making logic, particularly when it does not produce the expected results. Here we describe a new and enhanced visual interactive analysis of embeddings and explore its application in bioacoustics. Embeddings are the output of the penultimate layer of a DNN, an N-dimensional vector that, only one step removed from the final output, represent the inner-workings of a DNN model. Using existing dimensionality reduction techniques we converted the N-dimensional embeddings into 2 or 3-dimensional arrays displayed in scatterplots. By incorporating sound samples into the scatterplots we developed a visual and aural interactive interface and demonstrate its utility in assessing the performance of trained bioacoustic models, facilitating post-processing of results, error detection, input selection and the detection of rare events, which the reader can experience in online examples with publicly available code. • Deep Neural Networks are popular to process and identify bioacoustic sounds. • Scatterplots from processed embeddings reveal inner workings of DNN models. • Adding sound samples to embedding scatterplots facilitates their assessment. • We present a visual and aural interactive interface for assessing bioacoustic models. [ABSTRACT FROM AUTHOR]

The study of bird populations is crucial for biodiversity research and conservation. Deep artificial neural networks have revolutionized bird acoustic recognition; however, most methods overlook inherent relationships among bird populations, resulting in the loss of biological information. To address this limitation, we propose the Phylogenetic Perspective Neural Network (PPNN), which incorporates hierarchical multilevel labels for each bird. PPNN uses a hierarchical semantic embedding framework to capture feature information at different levels. Attention mechanisms are employed to extract and select common and distinguishing features, thereby improving classification accuracy. We also propose a path correction strategy to rectify inconsistent predictions. Experimental results on bird acoustic datasets demonstrate that PPNN outperforms current methods, achieving classification accuracies of 90.450%, 91.883%, and 89.950% on the Lishui-Zhejiang Birdsdata (100 species), BirdCLEF2018-Small (150 species), and BirdCLEF2018-Large (500 species) datasets respectively, with the lowest hierarchical distance of a mistake across all datasets. Our proposed method is applicable to any bird acoustic dataset and presents significant advantages as the number of categories increases. • A Phylogenetic Perspective Neural Network (PPNN) is developed for precise bird acoustic identification. • PPNN considers the hierarchical relationships among aves to enhance semantic coherence and reduce the severity of errors. • Position attention emphasizes shared features, while hierarchical attention identifies unique positions. [ABSTRACT FROM AUTHOR]

Continuous bioacoustic monitoring is an emerging opportunity as well as a challenge, allowing detection of cryptic species' activity while producing high computational demands. In this paper, we present an automated framework that allows the monitoring of a large number of bird species by their vocalizations over extended periods. The framework relies on the BirdNET-Analyzer deep learning model. We applied the framework to >80 species; 20 species with the highest recall scores were selected for further analysis. We used the framework to analyze acoustic signals recorded continuously for over two years using autonomous recorders at various locations in Agmon Hula Lake Park, Israel. During this period there was an acute outbreak of avian influenza in the area. We analyzed differences in acoustic occupancy for various species between two consecutive years (November 2020 to October 2022). We examined between-year population trends for 17 species, both migratory and resident, and found a significant decline in vocal activity between the two years for 10 species. We assume that this decline is related to the avian influenza outbreak, suggesting that the impact of the pandemic may be more widespread and affected a greater number of local species than was previously realized. This further highlights the power and effectiveness of bioacoustic monitoring in detecting cryptic but dramatic dynamics. [Display omitted] • Monitoring bioacoustics over two consecutive years, including an avian-flu outbreak. • An automated framework for tracking bird species occupancy is presented. • 'Vocalogram' - A visualization method of bioacoustic data is proposed. • Yearly comparisons show a significant decline in vocal activity for most species. • The decline in vocal activity may be linked to the avian-flu outbreak. [ABSTRACT FROM AUTHOR]

Acoustic sensors that collect acoustic data over extended periods and broad ranges are widely used in bioacoustics monitoring. However, in open environments, acoustic data collected using acoustic sensors can be subject to interference from various real-world noises, thereby influencing the subsequent analysis and processing of bioacoustic data. Existing bioacoustic noise reduction methods are limited in their application because of their low efficiency, unsuitability for non-stationary noise, generally unimproved signal-to-noise ratio (SNR) efficacy, and considerable amounts of residual noise. These limitations hinder the effective processing of recorded signals for which extraneous noise overlaps with bird vocalizations. In this study, we propose a bioacoustic noise reduction method based on a deep feature loss network for bird sounds. The method has a rapid denoising speed and can more effectively remove background noise from field recording signals without distorting the bird acoustic spectrum. The denoising effects of the proposed method were compared with those of a speech enhancement generative adversarial network, web real-time communications denoising, and other noise reduction methods. The denoising ability of these methods for different noises was evaluated using spectrograms and objective evaluation measures such as the SNR and perceptual evaluation of speech quality (PESQ). The experimental results revealed that our proposed noise reduction method can obtain higher SNRs and PESQ scores than other noise reduction methods, with the SNR increasing by up to 35.83 dB following denoising. • In open environments, acoustic data collected using acoustic sensor can be polluted. • Existing bioacoustics noise reduction methods are often limited. • We propose a bioacoustics noise reduction method for bird sounds. • Our method has a faster denoising speed and is more effective. • This noise reduction method can obtain higher SNR and PESQ. [ABSTRACT FROM AUTHOR]

Passive acoustic monitoring is increasingly used for assessing populations of marine mammals; however, analysis of large datasets is limited by our ability to easily classify sounds detected. Classification of beaked whale acoustic events, in particular, requires evaluation of multiple lines of evidence by expert analysts. Here we present a highly automated approach to acoustic detection and classification using supervised machine learning and open source software methods. Data from four large scale surveys of beaked whales (northwestern North Atlantic, southwestern North Atlantic, Hawaii, and eastern North Pacific) were analyzed using PAMGuard (acoustic detection), PAMpal (acoustic analysis) and BANTER (hierarchical random forest classifier). Overall classification accuracy ranged from 88% for the southwestern North Atlantic data to 97% for the northwestern North Atlantic. Results for many species could likely be improved with increased sample sizes, consideration of alternative automated detectors, and addition of relevant environmental features. These methods provide a highly automated approach to acoustic detection and classification using open source methods that can be readily adopted for other species and geographic regions. • BANTER is an open source, accessible machine learning classifier for data with multiple call types • BANTER can use multiple detectors on single call types for acoustic event classification • BANTER can classify beaked whale species with relatively low sample sizes • BANTER is easy to use and can be easily adopted to address current classification needs [ABSTRACT FROM AUTHOR]

Eastern Indian Ocean pygmy blue (EIOPB) whales were localised and tracked through the Perth Canyon area utilising a passive acoustic tracking array. Vocalising animals were tracked using the Type II unit of the EIOPB whale song across recordings for an average of 109.3 min (95% CI ± 27 min), suggesting that animals move slowly through the array area. Comparisons of the distribution of animals between sample years revealed non-uniform distribution of animals suggesting that environmental variables may drive pygmy blue whale spatial distribution within foraging areas. Analysis of the movement patterns of acoustically tracked animals found that animals tracked between January and early March generally exhibited directional movement while singing animals tracked between March and May exhibited directional and milling movement patterns, indicating that they may be foraging, and thus that animals may sing between feeding bouts. Comparisons of visual and acoustic localisations found animals appeared to utilise similar areas around the passive acoustic receiver array on days where survey effort overlapped, though acoustic survey methods returned consistently higher numbers of detections. Further, acoustic and visual detections of animals within the same day were not within the same location highlighting the need for holistic studies of species habitat utilisation. The inclusion of in-situ environmental data collection in the design of future acoustic tracking studies would be valuable to understand whether there is a link between environmental factors at the local scale and the localised movement of whales around the listening area. • Whales were tracked around a passive acoustic array by localising their signals. • Early in the season animals exhibited more directional movement patterns. • Later in the season animals spent more time milling. • Animals may balance behavioural trade-offs by singing between foraging bouts. [ABSTRACT FROM AUTHOR]

The use of autonomous recording units (ARUs) for passive acoustic monitoring has recently gained a lot of importance. It is now widely used for scientific purposes and has been increasingly applied in the field of conservation biology. However, while ARUs may greatly increase cost-effectiveness in monitoring, they are not exempt from possible biases. In this study, we compared the performances of three different software in recognizing bi- and tri-syllabic bird songs and tested if their results varied significantly in a real setting. We focused on the interpretation of the relative abundance among habitats and the time of vocal activity, also using different numbers of training records to assess if the results of the software improved. Even with the quality measures of the recognizer being consistent with those of previously published studies, in several cases the results produced significantly different interpretations depending on the software used. While increasing the number of training records slightly improved the overlap in the estimates of activity among the different software, differences in the relative abundance of each song were present even with a fairly solid number of training records. Our results suggest that reliable biological interpretations might only be attained with a large number of training records and very high values of precision, recall, and the F-score. In fact, even with very high precision and F-scores, lower values for recall led not only to statistically significant differences, but also to different interpretations of relative abundance and time of activity between the two best-performing software. Thus, great caution must be taken in interpreting the results of bioacoustics studies that use automated recognition, and standards of minimum quality should be created for recognizers in scientific and technical studies. • Automatic bird song recognition can yield different results depending on software. • Repetability of automatic bird song recognition studies can be compromised. • Increasing number of training records can improve automatic bird song recognition. • High quality measures should be demanded for automatic bird song recognition studies. [ABSTRACT FROM AUTHOR]

Convolutional neural networks (CNNs) have the potential to enable a revolution in bioacoustics, allowing robust detection and classification of marine sound sources. As global Passive Acoustic Monitoring (PAM) datasets continue to expand it is critical we improve our confidence in the performance of models across different marine environments, if we are to exploit the full ecological value of information within the data. This work demonstrates the transferability of developed CNN models to new acoustic environments by using a pre-trained model developed for one location (West of Scotland, UK) and deploying it in a distinctly different soundscape (Gulf of Mexico, USA). In this work transfer learning is used to fine-tune an existing open-source 'small-scale' CNN, which detects odontocete tonal and broadband call types and vessel noise (operating between 0 and 48 kHz). The CNN is fine-tuned on training sets of differing sizes, from the unseen site, to understand the adaptability of a network to new marine acoustic environments. Fine-tuning with a small sample of site-specific data significantly improves the performance of the CNN in the new environment, across all classes. We demonstrate an improved performance in area-under-curve (AUC) score of 0.30, across four classes by fine-training with only 50 spectrograms per class, with a 5% improvement in accuracy between 50 frames and 500 frames. This work shows that only a small amount of site-specific data is needed to retrain a CNN, enabling researchers to harness the power of existing pre-trained models for their own datasets. The marine bioacoustic domain will benefit from a larger pool of global data for training large deep learning models, but we illustrate in this work that domain adaptation can be improved with limited site-specific exemplars. • Convolutional Neural Networks (CNNs) are a popular method for detecting acoustic signals in large marine passive acoustic datasets. • To be of use to researchers algorithms must be robust across soundscapes which possess differing signal types and ambient noise conditions. • We fine-tuned an existing model with 50 - 500 training frames from a new location, to tailor the base modelto a new soundscape. • We report a30% improvement in model performance within the new marine soundscape after fine-tuning. • We demonstrate the ability to adapt CNNs to new acoustic environments with minimal training data for discriminating marine sound sources. [ABSTRACT FROM AUTHOR]

Abstract Owing to major technological advances, bioacoustics has become a burgeoning field in ecological research worldwide. Autonomous passive acoustic recorders are becoming widely used to monitor aerial insectivorous bats, and automatic classifiers have emerged to aid researchers in the daunting task of analysing the resulting massive acoustic datasets. However, the scarcity of comprehensive reference call libraries still hampers their wider application in highly diverse tropical assemblages. Capitalizing on a unique acoustic dataset of >650,000 bat call sequences collected over a 3-year period in the Brazilian Amazon, the aims of this study were (a) to assess how pre-identified recordings of free-flying and hand-released bats could be used to train an automatic classification algorithm (random forest), and (b) to optimize acoustic analysis protocols by combining automatic classification with visual post-validation, whereby we evaluated the proportion of sound files to be post-validated for different thresholds of classification accuracy. Classifiers were trained at species or sonotype (group of species with similar calls) level. Random forest models confirmed the reliability of using calls of both free-flying and hand-released bats to train custom-built automatic classifiers. To achieve a general classification accuracy of ~85%, random forest had to be trained with at least 500 pulses per species/sonotype. For seven out of 20 sonotypes, the most abundant in our dataset, we obtained high classification accuracy (>90%). Adopting a desired accuracy probability threshold of 95% for the random forest classifier, we found that the percentage of sound files required for manual post-validation could be reduced by up to 75%, a significant saving in terms of workload. Combining automatic classification with manual ID through fully customizable classifiers implemented in open-source software as demonstrated here shows great potential to help overcome the acknowledged risks and biases associated with the sole reliance on automatic classification. Highlights • Combining automatic and manual identification optimizes bat call classification. • Tested classifiers had high predictive performance for seven out of 20 species/sonotypes. • Models trained with calls from free-flying bats generate conservative classifications. • Using accuracy thresholds of 95% reduces workload of manual post-validation by up to 75%. [ABSTRACT FROM AUTHOR]

Environmental noise and data sparsity present major challenges within the field of bioacoustics. The presence of noise degrades the analysis of audio and field recordings commonly containing large quantities of data with sparse vocalisation features. This work explores noise reduction (audio enhancement) techniques in the context of extremely sparse vocalisations (< 1% occurrence rates) of invasive mammalian and marsupial species, and the clear implications for other bioacoustics applications which face similar challenges. This work compares relevant noise reduction techniques and recommends a spectral subtraction approach. Spectral subtraction achieved a 42.1 dB improvement in signal to noise power (SnNR) and a 2.7 dB improvement in noise variance (SR). We also demonstrate reliable noise reduction at bandwidths up to 250 kHz and efficiency improvements compared to alternative methods. We explore the benefits of deep audio enhancement approaches demonstrating comparable noise reduction with improvements in transient noise reduction but also key limitations such as bandwidth, efficiency and data generation in bioacoustics applications. We identify how the contributions of this work can be applied within the broader context of bioacoustics. All data and code is publicly available at https://github.com/BenMcEwen1/Sparse-Noise-Reduction • Evaluation of traditional and deep-learning-based noise reduction techniques in the context of a bioacoustics application.. • Demonstration of benefits and limitations of deep-learning-based approaches.. • Demonstration of the superior performance of spectral subtraction.. • Ablation study demonstrating the effects of bandwidth, and data requirements on noise reduction. • Discussion of related challenges within the broader context of bioacoustics and for noise reduction methods. [ABSTRACT FROM AUTHOR]

Automatic detection and classification of animal sounds has many applications in biodiversity monitoring and animal behavior. In the past twenty years, the volume of digitised wildlife sound available has massively increased, and automatic classification through deep learning now shows strong results. However, bioacoustics is not a single task but a vast range of small-scale tasks (such as individual ID, call type, emotional indication) with wide variety in data characteristics, and most bioacoustic tasks do not come with strongly-labelled training data. The standard paradigm of supervised learning, focussed on a single large-scale dataset and/or a generic pre-trained algorithm, is insufficient. In this work we recast bioacoustic sound event detection within the AI framework of few-shot learning. We adapt this framework to sound event detection, such that a system can be given the annotated start/end times of as few as 5 events, and can then detect events in long-duration audio—even when the sound category was not known at the time of algorithm training. We introduce a collection of open datasets designed to strongly test a system's ability to perform few-shot sound event detections, and we present the results of a public contest to address the task. Our analysis shows that prototypical networks are a very common used strategy and they perform well when enhanced with adaptations for general characteristics of animal sounds. However, systems with high time resolution capabilities perform the best in this challenge. We demonstrate that widely-varying sound event durations are an important factor in performance, as well as non-stationarity, i.e. gradual changes in conditions throughout the duration of a recording. For fine-grained bioacoustic recognition tasks without massive annotated training data, our analysis demonstrate that few-shot sound event detection is a powerful new method, strongly outperforming traditional signal-processing detection methods in the fully automated scenario. • Machine learning is increasingly used to detect animal calls, song and other vocalisations. • Standard "supervised learning" is common, but this requires a new detector to be trained for each new detection task. • We re-frame the problem as "few-shot learning", learning across multiple tasks, with only 5 training examples in each. • We introduce a deep learning baseline algorithm for this problem, and analyse in detail the outcomes of a public challenge. • Algorithms made this way generalise very strongly: detecting events even from sound categories not known at the time of algorithm training. [ABSTRACT FROM AUTHOR]

[Display omitted] • Acoustic monitoring: key tool for biodiversity assessment and ecological studies. • Harnessing metadata enhances deep bioacoustics classifier. • Geographical prior shows promise in leveraging citizen databases for bioacoustics. • Multi-branch CNN: significant improvements in bird song classification. • Metadata enables a 63% decrease in false positives for gibbon call detection. Bioacoustics, the exploration of animal vocalizations and natural soundscapes, has emerged as a valuable tool for studying species within their habitats, particularly those that are challenging to observe. This approach has broadened the horizons of biodiversity assessment and ecological research. However, monitoring wildlife with acoustic recorders produces large volumes of data that can be labor-intensive to analyze. Deep learning has recently transformed many computational disciplines by enabling the automated processing of large and complex datasets and has gained attention within the bioacoustics community. Despite the revolutionary impact of deep learning on acoustic detection and classification, attaining both high detection accuracy and low false positive rates in bioacoustics remains a significant challenge. An intriguing yet unexplored avenue for enhancing deep learning in bioacoustics involves the utilization of contextual information, such as time and location, to discern animal vocalizations within acoustic recordings. As a first case study, a multi-branch Convolutional Neural Network (CNN) was developed to classify 22 different bird songs using spectrograms as a first input, and spatial metadata as a secondary input. A comparison was made to a baseline model with only spectrogram input. A geographical prior neural network was trained, separately, to estimate the probability of a species occurring at a given location. The output of this network was combined with the baseline CNN. As a second case study, temporal data and spectrograms were used as input to a multi-branch CNN for the detection of Hainan gibbon (Nomascus hainanus) calls, the world's rarest primate. Our findings demonstrate that adding metadata to the bird song classifier significantly improves classification performance, with the highest improvement achieved using the geographical prior model (F1-score of 87.78% compared to 61.02% for the baseline model). The multi-branch CNNs also proved efficient (F1-scores of 76.87% and 78.77%) and simpler to use than the geographical prior. In the second case study, our findings revealed a decrease in false positives by 63% (94% of the calls were detected) when the metadata was used by the multi-branch CNN, and an increase of 19% in gibbon detection. This study has uncovered an exciting new avenue for improving classifier performance in bioacoustics. The methodology described in this study can assist ecologists, wildlife management teams, and researchers in reducing the amount of time spent analyzing large acoustic datasets obtained from passive acoustic monitoring studies. Our approach can be adapted and applied to other calling species, and thus tailored to other use cases. [ABSTRACT FROM AUTHOR]

[Display omitted] • We apply Deep Neural for Passive Acoustic Monitoring, comparing different architectures. • We also compare Deep NNs with other methods like MLP, Knn and Encoder- Decoder based in LSTM • We focus on the detection of Fin Whales calls in large datasets (on the order of hundreds of hours). • The final network works with images of different sizes, and it is suitable for working in real time. In this paper, we present an application of Deep Neural Networks to the detection of Fin Whales (Balaenoptera physalus) pulses, also known as calls , from very long acoustic recordings. For the purpose of detection, acoustic signals are converted to images using a Fourier transform operation. Therefore, acoustic pulses become specific shapes. The detection of pulses and their seasonal distribution is used by biologists to estimate the presence of animals and understand their behavior, and they have important ecological value. However, the variations in shape and the presence of background noise make detection difficult. The use of automated instruments for this task is crucial to processing the huge amount of data that comes from hundreds of hours of recordings in a fast and effective way. In this work, different network architectures are compared to assess their effectiveness in the task on a subset of recordings. Additionally, these are compared with other Machine Learning methods from the literature. The best performing network architecture is then applied to the whole set, consisting of three months of recordings, showing an accuracy of 81.37 % and a precision of 75 % , comparable to the results obtained with other methods. [ABSTRACT FROM AUTHOR]

The application of machine learning methods has led to major advances in the development of automated recognisers used to analyse bioacoustics data. To further improve the performance of automated call recognisers, we investigated the development of efficient data annotation strategies and how best to address uncertainty around ambiguous vocalisations. These challenges present a particular problem for species whose vocalisations are rare in field recordings, where collecting enough training data can be problematic and a species' vocalisations may be poorly documented. We provide an open access solution to address these challenges using two strategies. First, we applied an active learning framework to iteratively improve a convolutional neural network (CNN) model able to automate call identification for a target rare bird species, the southern black-throated finch (Poephila cincta cincta). We collected 9098 h of unlabelled audio recordings from a field study in the Desert Uplands Bioregion of Queensland, Australia, and used active learning to prioritise human annotation effort towards data that would best improve model fit. Second, we progressed methods for managing ambiguous vocalisations by applying machine learning methods more commonly used in medical image analysis and natural language processing. Specifically, we assessed agreement among human annotators and the CNN model (i.e. inter-annotator agreement) and used this to determine realistic performance outcomes for the CNN model and to identify areas where inter-annotator agreement may be improved. We also applied a classification approach that allowed the CNN model to classify sounds into an 'uncertain' category, which replicated a requirement of human-annotation and facilitated the comparison of human-model annotation performance. We found that active learning was an efficient strategy to build a CNN model where there was limited labelled training data available, and target calls were extremely rare in the unlabelled data. As few as five active learning iterations, generating a final labelled dataset of 1073 target calls and 5786 non-target sounds, were required to train a model to identify the target species with comparable performance to experts in the field. Assessment of inter-annotator agreement identified a bias in our model to align predictions most closely with those of the primary annotator and identified significant differences in inter-annotator agreement among subsets of our acoustic data. Our results highlight the use of inter-annotator agreement to understand model performance and identify areas for improvement in data annotation. We also show that excluding ambiguous vocalisations during data annotation results in an overestimation of model performance, an important consideration for datasets with inter-annotator disagreement. • Active learning is an efficient data annotation strategy for rare calls. • Assessing inter-annotator agreement benefits model evaluation and bias detection. • Excluding ambiguous vocalisations artificially inflates model performance. • We demonstrate an approach to manage ambiguous vocalisations in model development. [ABSTRACT FROM AUTHOR]

Non-invasive bioacoustic monitoring is becoming increasingly popular for biodiversity conservation. Two automated methods for acoustic classification of bird species currently used are frame-based methods, a model that uses Hidden Markov Models (HMMs), and event-based methods, a model consisting of descriptive measurements or restricted to tonal or harmonic vocalizations. In this work, we propose a new method for automated field recording analysis with improved automated segmentation and robust bird species classification. We used a Gaussian Mixture Model (GMM)-based frame selection with an event-energy-based sifting procedure that selected representative acoustic events. We employed a Mel, band-pass filter bank on each event's spectrogram. The output in each subband was parameterized by an autoregressive (AR) model, which resulted in a feature consisting of all model coefficients. Finally, a support vector machine (SVM) algorithm was used for classification. The significance of the proposed method lies in the parameterized features depicting the species-specific spectral pattern. This experiment used a control audio dataset and real-world audio dataset comprised of field recordings of eleven bird species from the Xeno-canto Archive, consisting of 2762 bird acoustic events with 339 detected “unknown” events (corresponding to noise or unknown species vocalizations). Compared with other recent approaches, our proposed method provides comparable identification performance with respect to the eleven species of interest. Meanwhile, superior robustness in real-world scenarios is achieved, which is expressed as the considerable improvement from 0.632 to 0.928 for the F-score metric regarding the “unknown” events. The advantage makes the proposed method more suitable for automated field recording analysis. [ABSTRACT FROM AUTHOR]

Many fish species use active sound production for communication in numerous behaviors. Additionally, likely all fish can make passive or incidental sounds that may also serve some signal functions. Despite the ecological importance of fish sounds, their evident passive acoustic monitoring applications, and extensive endeavors to document soniferous fish diversity, the fields of bioacoustics and ichthyology have historically lacked an easily accessible, global inventory of known fish sound production. To alleviate this limitation, we developed http://FishSounds.net , a website that compiles and disseminates fish sound production information and recordings. FishSounds Version 1.0 launched in 2021, cataloging documented examinations for active and passive sound production for 1185 fish species from 837 references as well as 239 exemplary audio recordings. FishSounds allows users to search by taxa (e.g., family or common name), geographical distribution (e.g., region or water body), sound type, or reference. We have also made available the code used to create the website, so that it may be used in other data-sharing efforts—acoustic or otherwise. Subsequent versions of the website will update the data and improve the website functionality. FishSounds will advance research into fish behavior, passive acoustic monitoring, and human impacts on underwater soundscapes; serve as a resource for public outreach; and provide the foundation needed to investigate more of the 96% of fish species that lack published examinations of sound production. We further hope the FishSounds design, implementation, and engagement strategies will serve as a model for future data management and sharing efforts. • The FishSounds website, launched in 2021, can be found at http://FishSounds.net. • We present sound information on 1185 fish species and 239 recordings. • Subsequent versions will update the data and improve website functionality. • The website code is available on GitLab for others to adapt and use. • FishSounds will aid research, outreach, management, and data sharing efforts. [ABSTRACT FROM AUTHOR]

Open audio databases such as Xeno-Canto are widely used to build datasets to explore bird song repertoire or to train models for automatic bird sound classification by deep learning algorithms. However, such databases suffer from the fact that bird sounds are weakly labelled: a species name is attributed to each audio recording without timestamps that provide the temporal localization of the bird song of interest. Manual annotations can solve this issue, but they are time consuming, expert-dependent, and cannot run on large datasets. Another solution consists in using a labelling function that automatically segments audio recordings before assigning a label to each segmented audio sample. Although labelling functions were introduced to expedite strong label assignment, their classification performance remains mostly unknown. To address this issue and reduce label noise (wrong label assignment) in large bird song datasets, we introduce a data-centric novel labelling function composed of three successive steps: 1) time-frequency sound unit segmentation, 2) feature computation for each sound unit, and 3) classification of each sound unit as bird song or noise with either an unsupervised DBSCAN algorithm or the supervised BirdNET neural network. The labelling function was optimized, validated, and tested on the songs of 44 West-Palearctic common bird species. We first showed that the segmentation of bird songs alone aggregated from 10% to 83% of label noise depending on the species. We also demonstrated that our labelling function was able to significantly reduce the initial label noise present in the dataset by up to a factor of three. Finally, we discuss different opportunities to design suitable labelling functions to build high-quality animal vocalizations with minimum expert annotation effort. • Open audio databases provide unique information for ecological sciences. • However, most of them are weakly labelled with a risk of signal confusion. • A labelling function based on segmentation, featurization, and clustering is proposed. • A test on the Xeno-Canto database revealed a significant reduction of label error. • Unsupervised labelling can foster the use of audio archives. [ABSTRACT FROM AUTHOR]

Bioacoustics has become widely used in the study of acoustically active animals, and machine learning algorithms have emerged as efficient and effective strategies to identify species vocalizations. Current applications of machine learning in bioacoustics often identify acoustic events to the species-level but fail to capture the complex acoustic repertoires animals use to communicate, which can inform habitat associations, demography, behavior, and the life history of cryptic species. The penultimate layer of most machine learning algorithms results in a vector of numbers describing the input, called feature embeddings. Here, we demonstrate that the feature embeddings generated by the BirdNET algorithm can enable within-species classifications of acoustic events. First, we successfully differentiated adult and juvenile Great Gray Owls; second, we identified three unique sounds associated with Great Spotted Woodpeckers (series call, alarm call, and drumming). These applications of BirdNET feature embeddings suggest that researchers can classify vocalizations into groups when group membership is unknown, and that within-species grouping is possible even when target signals are extremely rare. These applications of a relatively "black-box" aspect of machine learning algorithms can be used to derive ecologically informative acoustic classifications, which can inform the conservation of cryptic and otherwise difficult to study species. • AI tools can identify many species, but fail to represent animals' repertoires • BirdNET algorithm's feature embeddings enable within-species sound classification • Differentiating species' sounds can improve ecological and conservation research [ABSTRACT FROM AUTHOR]

Many studies have compared results from sound recordings and traditional point-count survey observer data when surveying avian communities. None have investigated the use of a moving sound recorder to replicate line-transect surveying. We conducted point-count surveys and line-transect surveys in four urban/peri-urban habitats in Darwin, tropical Australia, with stationary and moving sound recorders, respectively, to assess whether such a combination would result in more bird species being identified than with either technique alone. More bird species were identified using sound recordings than standard observer data. Further, the difference in the number of species identified between the observer and audio from point-count surveys was found to be significant with audio identification being more accurate; however, line-transect surveys showed no significant difference between the two identification methods. Overall, there was no statistical significance between using point-count surveys and line-transect surveys for total species identified. Linear mixed modelling found the interaction between habitat and survey type (point-count vs line-transect) was strongly significant, but not so that between habitat and survey method (sound recording vs human observation). Our results indicate that the integration of bioacoustic and drone technologies with traditional avian surveying techniques adds significant additional identifications when compiling a species list of an area. • The use of bioacoustic technology is advantageous when conducting avian surveys; • Drones can be a valuable tool when conducting avian surveys in challenging climatic conditions; • Bioacoustic and drone technology can be effectively combined to gather avian species data. [ABSTRACT FROM AUTHOR]

Passive acoustic monitoring coupled with automated signal recognition software has been widely used in recent years as an effective and affordable tool for wildlife monitoring and to combat illegal activities within protected areas. Here, we evaluate this technique to monitor the patterns of illegal cattle occurrence in the Brazilian Pantanal over a complete annual cycle. We aim to provide one of the first assessments of the performance of automated signal recognition software to detect ungulates. Cattle occurrences reached their maximum during the end of the dry season when lowland areas provide excellent pastures for cattle. In contrast, cattle occurrences were very low during the rainy season when the study area was seasonally inundated. Automated software was an efficient tool that was able to detect approximately three-quarters of cow calls within the recordings. Passive acoustic monitoring can be used to direct patrols to areas where illegal activities, such as cattle and poaching or logging, have been confirmed, which could be a method that would be especially well suited for remote areas, such as tropical forests. Future studies should evaluate whether there is a relationship between cattle grazing intensity and its associated impacts on wildlife and flora. Rapid advances in automated recognition and the recent development of low-cost recorders foresee a new era of acoustic ecology for improved conservation in the short term. • Cattle number is increasing and becoming an environmental issue. • Automated detection of cattle occurrence in protected areas is useful procedure. • Illegal cattle was detected in more than half of the monitored days. • Approximately 75% of cow calls were detected by automated software. • Passive acoustic monitoring may be used to direct patrols to conflict areas. [ABSTRACT FROM AUTHOR]

• In this purely data-driven approach we use machine learning models to perform classification of group-specific variations within the House Wren complex. • Additionally, we investigate geographical divergence in song characteristics within House Wren conspecifics by controlling for latitude. • We find out that our models are able to classify the variations with a high accuracy. • In order to provide reliable estimates of performance we employ a randomised approach to curate the dataset by drawing sound samples from different locations. • The results confirm an earlier field study of divergence in song characteristics in House Wren species. • The approach could be applied to other bird species in the database in order to discover or confirm variations in songs. Songbirds have shown variation in vocalizations across different populations and different geographical ranges. Such variations can over time lead to divergence in song characteristics, sometimes referred to as dialects. House Wren (Troglodytes aedon) is one such widely distributed bird species that has shown variation in its song characteristics within different populations. Traditionally, such studies have been conducted using manual approaches for classification. In this work we explore the use of machine learning models that can assist in performing classification of bird songs at a conspecific level. Two machine learning techniques, the random forest and a shallow feed forward neural network, are fed with pre-computed sound features to classify vocal variation in House Wren species across different reported population groups and latitudinal areas. A randomized approach is employed to create balanced subsets of sounds from different locations for repeated classification runs in order to provide a reliable estimate of performance. It is observed that such an automated approach is able to classify variations in songs within House Wren with high accuracy. We were also able to confirm the latitudinal variation of House Wren songs reported in previous studies. Given these results, we believe, such a purely data-driven way of analyzing bird songs in general can provide useful hints to biologists on where to look for interesting patterns in order to understand the evolutionary divergence in song characteristics. [ABSTRACT FROM AUTHOR]

• Rookognise, a deep learning system to acoustically identify individual animals. • Performance tested on a group of rooks, a social species with a rich repertoire. • Successful acoustic individual identification, whatever the vocalisation. • Designed for easy deployment for other species or additional classification tasks. Individual-level monitoring is essential in many behavioural and bioacoustics studies. Collecting and annotating those data is costly in terms of human effort, but necessary prior to conducting analysis. In particular, many studies on bird vocalisations also involve manipulating the animals or human presence during observations, which may bias vocal production. Autonomous recording units can be used to collect large amounts of data without human supervision, largely removing those sources of bias. Deep learning can further facilitate the annotation of large amounts of data, for instance to detect vocalisations, identify the species, or recognise the vocalisation types in recordings. Acoustic individual identification, however, has so far largely remained limited to a single vocalisation type for a given species. This has limited the use of those techniques for automated data collection on raw recordings, where many individuals can produce vocalisations of varying complexity, potentially overlapping one another, with the additional presence of unknown and varying background noise. This paper aims at bridging this gap by developing a system to identify individual animals in those difficult conditions. Our system leverages a combination of multi-scale information integration, multi-channel audio and multi-task learning. The multi-task learning paradigm is based the overall task into four sub-tasks, three of which are auxiliary tasks: the detection and segmentation of vocalisations against other noises, the classification of individuals vocalising at any point during a sample, and the sexing of detected vocalisations. The fourth task is the overall identification of individuals. To test our approach, we recorded a captive group of rooks, a Eurasian social corvid with a diverse vocal repertoire. We used a multi-microphone array and collected a large scale dataset of time-stamped and identified vocalisations recorded, and found the system to work reliably for the defined tasks. To our knowledge, the system is the first to acoustically identify individuals regardless of the vocalisation produced. Our system can readily assist data collection and individual monitoring of groups of animals in both outdoor and indoor settings, even across long periods of time, and regardless of a species' vocal complexity. All data and code used in this article is available online. [ABSTRACT FROM AUTHOR]

Classification of birdsong recordings can be naturally formulated as a multiple instance problem, where bags of instances are represented by either features or dissimilarities. In bioacoustics, bags typically correspond to regions of interest in spectrograms, which are detected after a segmentation stage of the audio recordings. In this paper, we use different dissimilarity measures between bags and explore whether the subsequent application of metric learning/adaptation methods and the construction of dissimilarity spaces allow increasing the classification performance of birdsong recordings. A publicly available bioacoustic data set is used for the experiments. Our results suggest, in the first place, that appropriate dissimilarity measures are those which capture most of the overall differences between bags, such as the modified Hausdorff distance and the mean minimum distance; in the second place, they confirm the benefit from adapting the applied dissimilarity measure as well as the potential further enhancement of the classification performance by building dissimilarity spaces and increasing training set sizes. [ABSTRACT FROM AUTHOR]

Recent expansion in the capabilities of passive acoustic monitoring of sound-producing animals is providing expansive data sets in many locations. These long-term data sets will allow the investigation of questions related to the ecology of sound-producing animals on time scales ranging from diel and seasonal to inter-annual and decadal. Analyses of these data often span multiple analysts from various research groups over several years of effort and, as a consequence, have begun to generate large amounts of scattered acoustic metadata. It has therefore become imperative to standardize the types of metadata being generated. A critical aspect of being able to learn from such large and varied acoustic data sets is providing consistent and transparent access that can enable the integration of various analysis efforts. This is juxtaposed with the need to include new information for specific research questions that evolve over time. Hence, a method is proposed for organizing acoustic metadata that addresses many of the problems associated with the retention of metadata from large passive acoustic data sets. A structure was developed for organizing acoustic metadata in a consistent manner, specifying required and optional terms to describe acoustic information derived from a recording. A client-server database was created to implement this data representation as a networked data service that can be accessed from several programming languages. Support for data import from a wide variety of sources such as spreadsheets and databases is provided. The implementation was extended to access Internet-available data products, permitting access to a variety of environmental information types (e.g. sea surface temperature, sunrise/sunset, etc.) from a wide range of sources as if they were part of the data service. This metadata service is in use at several institutions and has been used to track and analyze millions of acoustic detections from marine mammals, fish, elephants, and anthropogenic sound sources. [ABSTRACT FROM AUTHOR]

• We focused on data scarcity issues related to convolutional neural networks (CNNs). • Twelve modern pre-trained CNNs were compared. • This approach is less complex, and can be easily be adopted. • We contribute four passive acoustic datasets, corresponding to 90 h. • We achieved up to 82% F1 score on as few as 25 verified calls. Progress in deep learning, more specifically in using convolutional neural networks (CNNs) for the creation of classification models, has been tremendous in recent years. Within bioacoustics research, there has been a large number of recent studies that use CNNs. Designing CNN architectures from scratch is non-trivial and requires knowledge of machine learning. Furthermore, hyper-parameter tuning associated with CNNs is extremely time consuming and requires expensive hardware. In this paper we assess whether it is possible to build good bioacoustic classifiers by adapting and re-using existing CNNs pre-trained on the ImageNet dataset – instead of designing them from scratch, a strategy known as transfer learning that has proved highly successful in other domains. This study is a first attempt to conduct a large-scale investigation on how transfer learning can be used for passive acoustic monitoring (PAM), to simplify the implementation of CNNs and the design decisions when creating them, and to remove time consuming hyper-parameter tuning phases. We compare 12 modern CNN architectures across 4 passive acoustic datasets that target calls of the Hainan gibbon Nomascus hainanus , the critically endangered black-and-white ruffed lemur Varecia variegata , the vulnerable Thyolo alethe Chamaetylas choloensis , and the Pin-tailed whydah Vidua macroura. We focus our work on data scarcity issues by training PAM binary classification models very small datasets, with as few as 25 verified examples. Our findings reveal that transfer learning can result in up to 82% F1 score while keeping CNN implementation details to a minimum, thus rendering this approach accessible, easier to design, and speeding up further vocalisation annotations to create PAM robust models. [ABSTRACT FROM AUTHOR]

Projects on the acoustic monitoring of animals in natural habitats generally face the problem of managing extensive amounts of data, both needed for – and produced by – observation or experimentation. While there are many publicly accessible databases for recordings themselves, we are aware of none for annotated song sequences. In this paper, we describe our database system of bird vocalizations and introduce our online sample repository for the community of researchers studying the syntax of bird song. [ABSTRACT FROM AUTHOR]

Acoustic sensors allow scientists to scale environmental monitoring over large spatiotemporal scales. The faunal vocalisations captured by these sensors can answer ecological questions, however, identifying these vocalisations within recorded audio is difficult: automatic recognition is currently intractable and manual recognition is slow and error prone. In this paper, a semi-automated approach to call recognition is presented. An automated decision support tool is tested that assists users in the manual annotation process. The respective strengths of human and computer analysis are used to complement one another. The tool recommends the species of an unknown vocalisation and thereby minimises the need for the memorization of a large corpus of vocalisations. In the case of a folksonomic tagging system, recommending species tags also minimises the proliferation of redundant tag categories. We describe two algorithms: (1) a “naïve” decision support tool (16%–64% sensitivity) with efficiency of O ( n ) but which becomes unscalable as more data is added and (2) a scalable alternative with 48% sensitivity and an efficiency of O (log n ). The improved algorithm was also tested in a HTML-based annotation prototype. The result of this work is a decision support tool for annotating faunal acoustic events that may be utilised by other bioacoustics projects. [ABSTRACT FROM AUTHOR]

Ecologists are increasingly using bioacoustics in wildlife monitoring programs. Remote autonomous sound recorders provide new options for collecting data for species and in contexts that were previously difficult. However, post-processing of sound files to extract relevant data remains a significant challenge. Detection algorithms, or call recognizers, can aid automation of species detection but their performance and reliability has been mixed. Further, building recognizers typically requires either costly commercial software or expert programming skills, both of which reduces their accessibility to ecologists responsible for monitoring. In this study we investigated the performance of open-source call recognizers provided by the monitoR package in R, a language popular among ecologists. We tested recognizers on sound data collected under natural conditions at nests of two endangered subspecies of black-cockatoo, the Kangaroo Island glossy black-cockatoo Calyptorhynchus lathami halmaturinus (n = 23 nests), and the south-eastern red-tailed black-cockatoo Calyptorhynchus banksii graptogyne (n = 20 nests). Specifically, we tested the performance of binary point matching recognizers in confirming daily nest activity (active or inactive) and nesting outcome (fledge or fail). We tested recognizers on recordings from nests of known status using 3 × 3-h recordings per nest, from early, mid and late stages of the recording period. Daily nest activity was correctly assigned in 61.7% of survey days analysed (n = 60 days) for the red-tailed black-cockatoo, and 62.3% of survey days (n = 69 days) for the glossy black-cockatoo. Fledging was successfully detected in all cases. Precision (true positive / true positive + false positive) of individual detections was 70.2% for the south-eastern red-tailed black-cockatoo and 37.1% for the Kangaroo Island glossy black-cockatoo. Manual verification of outputs is still required, but it is not necessary to verify all detections to confirm an active nest (i.e., nest is deemed active when true positives are identified). We conclude that bioacoustics combined with semi-automated post-processing can be an appropriate tool for nest monitoring in these endangered subspecies. • Bioacoustic monitoring requires efficient and accessible call recognition methods. • We examined open-source recognizers for nest monitoring in black-cockatoos. • Recognizers correctly assigned nest activity in roughly 60% of survey days. • Fledging was detected by the recognizers. • Bioacoustics and semi-automated post-processing are useful for nest monitoring. [ABSTRACT FROM AUTHOR]

Monitoring of biological populations is well known for being a complex task that involves high operational costs, unknown reproductive intervals of the studied species, and difficult visualization of isolated individuals (due to their mimetic and cryptic capabilities). Therefore, the development of new methodologies able to measure quantities of individuals in specific biological populations without direct contact is desired. Species and individual recognition, based on acoustic analysis of their calls (Bioacoustics), is possible for many animals and has proven to be a useful tool in the study and monitoring of animal species. In this paper, an unsupervised methodology for anuran automatic identification is proposed; it is based on the use of a fuzzy classifier and Mel Frequency Cepstral Coefficients. This methodology is able to detect species not presented in the training stage, although they belong to different populations. Additionally, correlations among species of the same genus can be determined through the similarities of their calls. For testing the proposed method, two different datasets with species from the northeastern Colombia (Chocó and Antioquia departments with 103 and 813 mating calls respectively) were used. In validation tests performed, accuracies between 99.38% and 100% were achieved in all species by applying the proposed methodology to both datasets. Thirteen different species of anurans in both datasets were correctly identified. [ABSTRACT FROM AUTHOR]

Abstract: Passive acoustic monitoring often leads to large quantities of sound data which are burdensome to process, such that the availability and cost of expert human analysts can be a bottleneck and make ecosystem or landscape-scale projects infeasible. This manuscript presents a method for rapidly analyzing the results of band-limited energy detectors, which are commonly used for the detection of passerine nocturnal flight calls, but which typically are beset by high false positive rates. We first manually classify a subset of the detected events as signals of interest or false detections. From that subset, we build a Random Forest model to eliminate most of the remaining events as false detections without further human inspection. The overall reduction in the labor required to separate signals of interest from false detections can be 80% or more. Additionally, we present an R package, flightcallr, containing functions which can be used to implement this new workflow. [Copyright &y& Elsevier]

Abstract: Numerous methods are available for analysis of avian vocalizations, but few research efforts have compared recent methods for calculating and evaluating similarity among calls, particularly those collected in the field. This manuscript compares a suite of methodologies for analyzing flight calls of New World warblers, investigating the effectiveness of four techniques for calculating call similarity: (1) spectrographic cross-correlation, (2) dynamic time warping, (3) Euclidean distance between spectrogram-based feature measurements, and (4) random forest distance between spectrogram-based feature measurements. We tested these methods on flight calls, which are short, structurally simple vocalizations typically used during nocturnal migration, as these signals may contain important ecological or demographic information. Using the four techniques listed above, we classified flight calls from three datasets, one collected from captive birds and two collected from wild birds in the field. Each dataset contained an equal number of calls from four warbler species commonly recorded during acoustic monitoring: American Redstart, Chestnut-sided Warbler, Hooded Warbler, and Ovenbird. Using captive recordings to train the classification models, we created four similarity-based classifiers which were then tested on the captive and field datasets. We show that these classification methods are limited in their ability to successfully classify the calls of these warbler species, and that classification accuracy was lower on field recordings than captive recordings for each of the tested methods. Of the four methods we compared, the random forest technique had the highest classification accuracy, enabling correct classification of 67.6% of field recordings. To compare the performance of the automated techniques to manual classification, the most common method used in flight call research, human experts were also asked to classify calls from each dataset. The experts correctly classified approximately 90% of field recordings, indicating that although the automated techniques are faster, they remain less accurate than manual classification. However, because of the challenges inherent to these data, such as the structural similarity among the flight calls of focal species and the presence of environmental noise in the field recordings, some of the tested automated classification techniques may be acceptable for real-world applications. We believe that this comparison of broadly applicable methodologies provides information that will prove to be useful for analysis, detection and classification of short duration signals. Based on our results, we recommend that a combination of feature measurements and random forest classification can be used to assign flight calls to species, while human experts oversee the process. [Copyright &y& Elsevier]

Abstract: Increasing concern about decline in biodiversity has created a demand for population surveys. Acoustic monitoring is an efficient non-invasive method, which may be deployed for surveys of animals as diverse as insects, birds, and bats. Long-term unmanned automatic monitoring may provide unique unbiased data from a whole season, but the large amount of data presents serious challenges for the automatic processing of the measurements. To demonstrate feasibility of automatic multi-channel surveying using a new prototype hardware, we carried out a 2-month study of echolocating bats requiring high data sampling rates (500kHz). Using a sound energy threshold criterion for triggering recording, we collected 236GB (Gi=10243) of data at full bandwidth. We implemented a simple automatic method using a Support Vector Machine (SVM) classifier based on a combination of temporal and spectral analyses to classify events into bat calls and non-bat events. After experimentation we selected duration, energy, bandwidth, and entropy as classification features to identify short high energy structured sounds in the right frequency range. The spectral entropy makes use of the orderly arrangement of frequencies in bat calls to reject short noise pulses, e.g. from rain. The SVM classifier reduced our dataset to 162MB of candidate bat calls with an estimated accuracy of 96% for dry nights and 70% when it was raining. The automatic survey revealed calls from two species of bat not previously recorded in the area, as well as an unexpected abundance of social calls. The recordings provide data which can be used to correlate bat activity with rain, temperature, and sunset/sunrise. We discuss future applications, achieving higher accuracy in classifying bat calls and the possibility of using trajectory-tracking data to determine bat behavior and correct for the bias toward loud bats inherent in acoustic surveying. [Copyright &y& Elsevier]

Abstract: Bioacoustic monitoring is becoming more and more popular as a non-invasive method to study populations and communities of vocalizing animals. Acoustic pattern recognition techniques allow for automated identification of species and an estimation of species composition within ecosystems. Here we describe an approach where on the basis of long term acoustic recordings not only the occurrence of a species was documented, but where the number of vocalizing animals was also estimated. This approach allows us to follow up changes in population density and to define breeding sites in a changing environment. We present the results of five years of continuous acoustic monitoring of Eurasian bittern (Botaurus stellaris) in a recent wetland restoration area. Using a setup consisting of four four-channel recorders equipped with cardioid microphones we recorded vocal activity during entire nights. Vocalizations of bitterns were detected on the recordings by spectrogram template matching. On basis of time differences of arrival (TDOA) of the acoustic signals at different recording devices booming bitterns could be mapped using hyperbolic localization. During the study period not only changes in the number of calling birds but also changes in their spatial distribution connected with changes in habitat structure could be documented. This semi-automated approach towards monitoring birds described here could be applied to a wide range of monitoring tasks for animals with long distance vocalizations. [Copyright &y& Elsevier]

One quarter of all terrestrial native bird species have become extinct since human arrival in New Zealand, leading to a pervasive silence in many natural environments due to the decrease in native bird song. Passive acoustic techniques are a potential tool for environmental monitoring, especially for testing whether the control of mammals can reverse the 'silent forest' effect. Here we compare soundscapes from two nearby sites within the Waitakere Ranges Regional Park, New Zealand, that have contrasting predator control levels: one with high-level pest mammal control, and the other with low-level pest control. Measurements of twelve acoustic indices extracted from two seasons of passive acoustic recordings are split into 20 acoustic regions to identify which regions best discriminate between the two management regimes. We define the acoustic regions as units of analysis bounded by a specific time period and frequency range chosen to capture the main groups of biologically relevant acoustic events within a soundscape. Analysis of variance and pairwise comparisons indicated the acoustic region bounded from 9 pm to 11:59 pm and a range of 0.988–3.609 kHz in autumn presented the greatest differences between sites. The sounds responsible for these acoustic differences were generated by invasive mammals in the site with no pest control. Results also supports spring season as the most important for bird monitoring in New Zealand. Acoustic indices analysis did not detect a reversal of the "silence forest" effect in the site with high-level predator control. • Acoustic indices split into acoustic regions facilitate soundscape comparisons. • Acoustic regions are defined by frequency and time ranges. • Pest control did not reverse the silent forest effect in the Waitakere. • Acoustic regions enable detecting soundscape differences. [ABSTRACT FROM AUTHOR]

Categorizing the bioacoustic and ecoacoustic properties of animals is great interest to biologists and ecologists. Also, multidisciplinary studies in engineering have significantly contributed to the development of acoustic analysis. Observing the animals living in the ecological environment provides information in many areas such as global warming, climate changes, monitoring of endangered animals, agricultural activities. However, the classification of bioacoustics sounds by manually is very hard. Therefore, automated bioacoustics sound classification is crucial for ecological science. This work presents a new multispecies bioacoustics sound dataset and novel machine learning model to classify bird and anuran species with sounds automatically. In this model, a new nonlinear textural feature generation function is presented by using twine cipher substitution box(S-box), and this feature generation function is named twine-pat. By using twine-pat and tunable Q-factor wavelet transform, a multilevel feature generation network is presented. Iterative ReliefF(IRF) is employed to select the most effective/valuable features. Two shallow classifiers are used to calculate results. Our presented model reached 98.75% accuracy by using k-nearest neighbor(kNN) classifier. The results obviously demonstrated the success of the presented model. • A new multispecies sound dataset was collected. • A nonlinear feature generation function was proposed to show the feature extraction ability of the S-box. • A high accurate sound classification model was presented by using the proposed twine-pat iterative ReliefF. • This method outperformed. [ABSTRACT FROM AUTHOR]

Abstract: Today''s acoustic monitoring devices are capable of recording and storing tremendous amounts of data. Until recently, the classification of animal vocalizations from field recordings has been relegated to qualitative approaches. For large-scale acoustic monitoring studies, qualitative approaches are very time-consuming and suffer from the bias of subjectivity. Recent developments in supervised learning techniques can provide rapid, accurate, species-level classification of bioacoustics data. We compared the classification performances of four supervised learning techniques (random forests, support vector machines, artificial neural networks, and discriminant function analysis) for five different classification tasks using bat echolocation calls recorded by a popular frequency-division bat detector. We found that all classifiers performed similarly in terms of overall accuracy with the exception of discriminant function analysis, which had the lowest average performance metrics. Random forests had the advantage of high sensitivities, specificities, and predictive powers across the majority of classification tasks, and also provided metrics for determining the relative importance of call features in distinguishing between groups. Overall classification accuracy for each task was slightly lower than reported accuracies using calls recorded by time-expansion detectors. Myotis spp. were particularly difficult to separate; classifiers performed best when members of this genus were combined in genus-level classification and analyzed separately at the level of species. Additionally, we identified and ranked the relative contributions of all predictor features to classifier accuracy and found measurements of frequency, total call duration, and characteristic slope to be the most important contributors to classification success. We provide recommendations to maximize accuracy and efficiency when analyzing acoustic data, and suggest an application of automated bioacoustics monitoring to contribute to wildlife monitoring efforts. [Copyright &y& Elsevier]

• The algorithm reached 33% and 5% F1 for the classes drumming and vocalization, respectively, improving in previous results. • Spectrogram denoising operations significantly improved detection performance. • Networks performed best when trained on the heavily imbalanced training data set, i.e. without resampling. • Standard binary cross entropy outperformed alternative loss variants that are specialized in compensating class imbalance. Automatic detection systems are important in passive acoustic monitoring (PAM) systems, as these record large amounts of audio data which are infeasible for humans to evaluate manually. In this paper we evaluated methods for compensating class imbalance for deep-learning based automatic detection of acoustic chimpanzee calls. The prevalence of chimpanzee calls in natural habitats is very rare, i.e. databases feature a heavy imbalance between background and target calls. Such imbalances can have negative effects on classifier performances. We employed a state-of-the-art detection approach based on convolutional recurrent neural networks (CRNNs). We extended the detection pipeline through various stages for compensating class imbalance. These included (1) spectrogram denoising, (2) alternative loss functions, and (3) resampling. Our key findings are: (1) spectrogram denoising operations significantly improved performance for both target classes, (2) standard binary cross entropy reached the highest performance, and (3) manipulating relative class imbalance through resampling either decreased or maintained performance depending on the target class. Finally, we reached detection performances of 33% F 1 for drumming and 5% F 1 for vocalization, which is a >7 fold increase compared to previously published results. We conclude that supporting the network to learn decoupling noise conditions from foreground classes is of primary importance for increasing performance. [ABSTRACT FROM AUTHOR]

In the study of animal behaviour, annotation and analysis is largely done manually either directly in the field or from recordings. An emerging field, computational ethology, is challenging this approach by using machine learning to automate the process. However, the use of such methods in general is complicated by a lack of modularity, leading to high cost and long development times. At the same time, the benefits of implementing a fully automated pipeline are often minuscule. We propose online analysis as a way to gain more from automating the process, such as making it easier to ensure that equipment is properly configured and calibrated, enabling the recording equipment to follow the animals, and even enabling closed-loop experiments. In this work, we discuss the requirements and challenges for such a system and propose an implementation based on modern IT infrastructure. Finally, we demonstrate the system in case studies of bats and mongoose. As more and more methods and algorithms are developed we expect online systems to enable new experimental setups to study behaviour, leading to new insights in the field. [ABSTRACT FROM AUTHOR]

Variation in avian diversity in space and time is commonly used as a metric to assess environmental changes. Conventionally, such data were collected by expert observers, but passively collected acoustic data is rapidly emerging as an alternative survey technique. However, efficiently extracting accurate species richness data from large audio datasets has proven challenging. Recent advances in deep artificial neural networks (DNNs) have transformed the field of machine learning, frequently outperforming traditional signal processing techniques in the domain of acoustic event detection and classification. We developed a DNN, called BirdNET, capable of identifying 984 North American and European bird species by sound. Our task-specific model architecture was derived from the family of residual networks (ResNets), consisted of 157 layers with more than 27 million parameters, and was trained using extensive data pre-processing, augmentation, and mixup. We tested the model against three independent datasets: (a) 22,960 single-species recordings; (b) 286 h of fully annotated soundscape data collected by an array of autonomous recording units in a design analogous to what researchers might use to measure avian diversity in a field setting; and (c) 33,670 h of soundscape data from a single high-quality omnidirectional microphone deployed near four eBird hotspots frequented by expert birders. We found that domain-specific data augmentation is key to build models that are robust against high ambient noise levels and can cope with overlapping vocalizations. Task-specific model designs and training regimes for audio event recognition perform on-par with very complex architectures used in other domains (e.g., object detection in images). We also found that high temporal resolution of input spectrograms (short FFT window length) improves the classification performance for bird sounds. In summary, BirdNET achieved a mean average precision of 0.791 for single-species recordings, a F0.5 score of 0.414 for annotated soundscapes, and an average correlation of 0.251 with hotspot observation across 121 species and 4 years of audio data. By enabling the efficient extraction of the vocalizations of many hundreds of bird species from potentially vast amounts of audio data, BirdNET and similar tools have the potential to add tremendous value to existing and future passively collected audio datasets and may transform the field of avian ecology and conservation. • We developed a DNN, called BirdNET, capable of identifying 984 North American and European bird species by sound • We tested the model against three independent datasets of single-species recordings and fully annotated soundscape data • BirdNET achieved a mean average precision of 0.791 for single-species recordings and a F0.5 score of 0.414 for annotated soundscapes • Our system is publicly available and can be used by other researchers • We also provide online demos and prototypes on out project website at https://birdnet.cornell.edu [ABSTRACT FROM AUTHOR]

Hidden Markov Models (HMMs) were developed and implemented for the discrimination of 11 available Whales (Cetaceans). The primarily aims of the experiments were to explore the impact frame size and step size, feature vector size, and number of states for feature extraction and acoustic models on classification accuracy. Through the experiments using Mel-Frequency Cepstral Coefficients (MFCCs) extracted from the vocalizations (7 ms frame size and 6 ms step size), HMMs containing 4 states with single underlying Gaussian Mixture Model (GMM) yielded high classification accuracies ranging from 82.72% (9 classes) to 100.00% (1–3 classes), including discrimination of 84.11% (11 classes). From the results, the framework could be applied to the analysis of other marine mammals for the automatic classification and detection of vocalizations and species. [ABSTRACT FROM AUTHOR]