Hidden in the nook of a tree, a small microphone records the sounds of the forest throughout the night. Meanwhile, a bird flaps past with a small GPS tag mounted to its back. Hundreds of kilometres away, a computer algorithm scans the recordings and identifies the calls, while scientists track the bird’s movements from behind their screens. 

The tide of artificial intelligence is rising in biodiversity monitoring, where deep learning and artificial neural networks are used to process vast amounts of data, amplifying the speed and scale of scientific analysis. This is coupled with technological advances that allow biologists to collect data in novel ways and at a lower cost. Researchers can now embed relatively cheap wildlife sensors such as camera traps, Autonomous Recording Units (ARUs), and thermal cameras across the landscape, allowing them to monitor difficult-to-capture species, access hard-to-reach locations, and expand their study scope across time and space. 

Machine listening refers to the ability of computers to interpret sound. Recorders capture species calls and soundscapes and pass recordings through algorithms to answer biological and ecological questions. One example of the use of machine listening is to process long nighttime recordings of bird calls produced by nocturnal migrants. Up to 85% of birds migrate at night, and most have species-specific nocturnal flight calls [1]. 

In the past, people relied on techniques such as moon watching – observing the moon through a telescope to spot the outline of birds flying by [2]! Scientists and hobbyists can instead scroll through spectrograms during the day to search for potential calls [3]. Although faster than night-long activities, manually labelling datasets remains extremely labour-intensive. This is where machine listening is transforming the field, allowing computers to detect and identify calls and species in a matter of minutes.

But these systems are not perfect. On the ground, the precision and accuracy of machine learning models used to identify species and call types depends on how the models are trained. Trying to detect a European bird on a model trained on Hawaiian species might incorrectly identify a Kiwikiu rather than the far more likely Cuckoo. 

However, model development is moving forward rapidly; artificial neural networks can analyse spectrograms using computer vision, and feature embeddings can capture even the slightest of differences between sounds [4]. This approach allows models to be pre-trained and then fine-tuned with local data using transfer learning – far more accessible than building a model from scratch, not to mention less computationally expensive [5].

The spread of wildlife sensors raises questions beyond technology. In a recent publication titled “The Planetisation of Machine Listening”, Professor James E. K. Parker argues that the more machine listening is used as a monitoring and conservation tool, the more its use will expand [6]. Once we start – and we already have – we are inclined to upscale at a pace with which political and legal frameworks are not currently equipped to deal with. 

Unresolved challenges regarding carbon-friendly data storage and data security persist throughout the field, as do ethical considerations when bioacoustics begin to “actively interact” with nature, for example when loudspeakers play sounds in coral restoration zones to call in specific species [6].

Figure 1: “Camera Trap” by jurvetson is licensed under CC BY 2.0. 2009.