Periodic Reporting for period 1 - BioacAI (Bioacoustic AI for wildlife protection)
Reporting period: 2023-09-01 to 2025-08-31
Biodiversity loss is ranked as one of the top 5 risks globally, both in impact and likelihood. Anticipated significant declines in bird and insect populations could trigger profound ecological catastrophes. Yet data for monitoring biodiversity change remain incomplete. It is crucial to monitor the world's wildlife better, faster, and more completely.
Sound recording could help: it is cheap and rapid, and modern machine learning can radically improve its precision. However, there are two barriers: AI-enhanced acoustic tools for wildlife monitoring are mostly prototypes; and few experts combine knowledge of animals, sound, and AI.
The BioacAI project addresses these challenges by developing new AI methods directly within the context of acoustic wildlife monitoring - using new acoustic devices, algorithms, and insights into ecology and behaviour.
Sound recording could help: it is cheap and rapid, and modern machine learning can radically improve its precision. However, there are two barriers: AI-enhanced acoustic tools for wildlife monitoring are mostly prototypes; and few experts combine knowledge of animals, sound, and AI.
The BioacAI project addresses these challenges by developing new AI methods directly within the context of acoustic wildlife monitoring - using new acoustic devices, algorithms, and insights into ecology and behaviour.
BioacAI develops new AI methods directly within the context of acoustic wildlife monitoring. It has a cross-disciplinary span, which can be divided into four themes: AI methods, on-device algorithms, behaviour analysis and ecological monitoring. The project is structured to progress on all four of these in parallel, and to make connections between them. Progress on all four is proceeding as planned.
Hardware and devices: Some parts of the project use existing devices, such as the well-known Audiomoth sound recorder. Others have been developing new or refined hardware designs: in particular, designs and strategies for 3D spatial analysis of sound patterns, through techniques such as beamforming and sound source localisation, using multi-microphone “arrays”. Prototypes have spanned a wide spectrum of hardware types from small, distributed microphone arrays to large microphone arrays including 64 microphones. For two of the DCs working on spatial sound, a crucial step has been the design of simulation environments tailored for bioacoustics where spatial algorithms could be prototyped. We have also investigated the ecological footprint of bioacoustic hardware, eliciting requirements from practitioner interviews and making initial designs for low-footprint strategies which will be further developed and evaluated in the second half of the project.
The computational methods used in Bioacoustic AI come under two main categories: signal processing, and AI. A collaboration among all of the DCs resulted in the first contribution of a technical report, which gives an overview of on-device signal processing and artificial intelligence (AI) algorithms used in bioacoustics.
On the AI/machine learning side specifically, one focus is to move beyond basic classification, to tasks that suit the special constraints of wildlife monitoring tasks. To this end, initial results from various ML strategies have been developed, in three separate PhD projects. One uses human-in-the-loop “active learning”; one explores feature representations from pre-trained deep “embeddings” to facilitate discovery of rare sounds; and one has investigated how to use multiple datasets together in “multi-task” AI. The first results from these have been published in research workshops. Software implementations have also been made public, in particular a software framework “bacpipe” https://github.com/bioacoustic-ai/bacpipe(opens in new window) to provide the core task of deep feature extraction from sounds.
Animal sound datasets come from multiple sources. There are pre-existing datasets, though highly variable in their size and the detail of their metadata. To assist with navigating and curating these, we built an online web directory “Bioacoustics Datasets” https://bioacoustic-ai.github.io/bioacoustics-datasets/(opens in new window) which lists over 90 existing datasets representing many different types of animals. Additionally, we have curated data into new subsets, and/or add new manual annotations of phenomena. All of these are progressing satisfactorily. Open data publication will largely occur at the same time as the publication of the research papers describing them.
Animal behaviour is the focus of two DCs, using audio to characterise animal interactions and social information sharing. For the work on hyenas, existing audio data has been annotated and curated, and a recent AI model “animal2vec” has been customised. For the work on birds (yellowhammer), analysis of pre-collected datasets has revealed social effects of distance. Fieldwork has also been conducted to perform sound transmission experiments.
Ecological monitoring will benefit from the increased quality of acoustic surveying we develop. For birds, DC9 conducted sound transmission experiments in the field to measure how the identifiability of individual song varies with distance, and used this to design and then deploy Audiomoth recorders in the main fieldwork carried out in 2025 from March to July. For bats, DC10 curated data and trained an initial AI model for bat call detection, and also performed a wide range of community coordination/facilitation activities to build capacity and collaborations for large-scale bat recognition algorithms.
Hardware and devices: Some parts of the project use existing devices, such as the well-known Audiomoth sound recorder. Others have been developing new or refined hardware designs: in particular, designs and strategies for 3D spatial analysis of sound patterns, through techniques such as beamforming and sound source localisation, using multi-microphone “arrays”. Prototypes have spanned a wide spectrum of hardware types from small, distributed microphone arrays to large microphone arrays including 64 microphones. For two of the DCs working on spatial sound, a crucial step has been the design of simulation environments tailored for bioacoustics where spatial algorithms could be prototyped. We have also investigated the ecological footprint of bioacoustic hardware, eliciting requirements from practitioner interviews and making initial designs for low-footprint strategies which will be further developed and evaluated in the second half of the project.
The computational methods used in Bioacoustic AI come under two main categories: signal processing, and AI. A collaboration among all of the DCs resulted in the first contribution of a technical report, which gives an overview of on-device signal processing and artificial intelligence (AI) algorithms used in bioacoustics.
On the AI/machine learning side specifically, one focus is to move beyond basic classification, to tasks that suit the special constraints of wildlife monitoring tasks. To this end, initial results from various ML strategies have been developed, in three separate PhD projects. One uses human-in-the-loop “active learning”; one explores feature representations from pre-trained deep “embeddings” to facilitate discovery of rare sounds; and one has investigated how to use multiple datasets together in “multi-task” AI. The first results from these have been published in research workshops. Software implementations have also been made public, in particular a software framework “bacpipe” https://github.com/bioacoustic-ai/bacpipe(opens in new window) to provide the core task of deep feature extraction from sounds.
Animal sound datasets come from multiple sources. There are pre-existing datasets, though highly variable in their size and the detail of their metadata. To assist with navigating and curating these, we built an online web directory “Bioacoustics Datasets” https://bioacoustic-ai.github.io/bioacoustics-datasets/(opens in new window) which lists over 90 existing datasets representing many different types of animals. Additionally, we have curated data into new subsets, and/or add new manual annotations of phenomena. All of these are progressing satisfactorily. Open data publication will largely occur at the same time as the publication of the research papers describing them.
Animal behaviour is the focus of two DCs, using audio to characterise animal interactions and social information sharing. For the work on hyenas, existing audio data has been annotated and curated, and a recent AI model “animal2vec” has been customised. For the work on birds (yellowhammer), analysis of pre-collected datasets has revealed social effects of distance. Fieldwork has also been conducted to perform sound transmission experiments.
Ecological monitoring will benefit from the increased quality of acoustic surveying we develop. For birds, DC9 conducted sound transmission experiments in the field to measure how the identifiability of individual song varies with distance, and used this to design and then deploy Audiomoth recorders in the main fieldwork carried out in 2025 from March to July. For bats, DC10 curated data and trained an initial AI model for bat call detection, and also performed a wide range of community coordination/facilitation activities to build capacity and collaborations for large-scale bat recognition algorithms.
In the first phase of the project, the main novel results that are available thus far relate to signal processing and machine learning methods. These have been published openly in peer-reviewed conference venues such as the European Signal Processing Conference (EUSIPCO), Forum Acusticum, and the International Bioacoustics Congress (IBAC). These results provide improvements in the ways that algorithmic analysis can be used automatically to spatially locate animal sounds, and analyse their content. Separately, we have published strategic perspectives on how to lower the ecological footprint of automated monitoring.
Results on animal behaviour and ecology are expected to grow more strongly in the second half of the project, due to the fieldwork and other time investment required before first analysis. However, some early results are available, e.g. a collaboration using existing datasets of bird recordings in France to understand how environmental factors affect bird behaviour, published in the journal Bird Study.
To ensure further uptake and success, there are two primary needs:
Firstly, we need further ways to expand the skills base in the wider community - such as more opportunities for ecologists to learn, to network, and to stay up-to-date on the use of these AI methods, and for computer scientists to learn how they can contribute to ecology. The Doctoral Network itself provides a platform for this, but it is clear that the demand for such skills is much larger.
Secondly, we need large-scale support for sharing of digital material such as audio data or trained machine learning models. While initiatives such as GBIF, EOSC and Zenodo help to share animal observation metadata and research outputs, there is an unmet need for semi-open sharing of large volumes of media data (with individual data records often being more than 1TB). This does not refer only to research outputs but intermediate products such as training data. A European-level facility could be an ideal way to coordinate this.
Results on animal behaviour and ecology are expected to grow more strongly in the second half of the project, due to the fieldwork and other time investment required before first analysis. However, some early results are available, e.g. a collaboration using existing datasets of bird recordings in France to understand how environmental factors affect bird behaviour, published in the journal Bird Study.
To ensure further uptake and success, there are two primary needs:
Firstly, we need further ways to expand the skills base in the wider community - such as more opportunities for ecologists to learn, to network, and to stay up-to-date on the use of these AI methods, and for computer scientists to learn how they can contribute to ecology. The Doctoral Network itself provides a platform for this, but it is clear that the demand for such skills is much larger.
Secondly, we need large-scale support for sharing of digital material such as audio data or trained machine learning models. While initiatives such as GBIF, EOSC and Zenodo help to share animal observation metadata and research outputs, there is an unmet need for semi-open sharing of large volumes of media data (with individual data records often being more than 1TB). This does not refer only to research outputs but intermediate products such as training data. A European-level facility could be an ideal way to coordinate this.