Case Study:

Using Machine Learning To Identify Birdsongs

Summary

Artificial intelligence is learning how to identify birdcalls, enabling scientists to discover the location and number of New Zealand endemic birds. Knowing where the birds are and how many will assist biologists to better understand and conserve threatened species.

Contributor

Dr. Victor Anton

The Conservation Challenge

Researchers from Victoria University of Wellington, Victor Anton, Stephen Hartley, and Heiko Wittmer have recorded tens-of-thousands of birdcalls of currently threatened species such as Hihi (Stitchbird), Tīeke (Saddleback), and Kākāriki (Red-crowned parakeet). The audio was recorded using acoustic sensors at 50 locations in and around Zealandia – New Zealand’s first fenced eco-sanctuary and world-class conservation project. Identifying the location and number of different birdcalls in the audio enables biologists to identify factors influencing the threatened birds outside the eco-sanctuary. However, the prospect of sitting down to analyse so much audio threatened to overwhelm the researchers.

New Zealand endangered bird species, , Victor Anton
The three New Zealand bird species targeted in the study: Kākāriki, Hihi, and Tīeke

The Solution

With the help of Paul Matthews (artificial intelligence expert at NEC New Zealand), researchers are developing birdsong recognition technology to do the job faster and more accurately.

Description of Technology Used

Using the deep neural network algorithms, the team is working on a system that learns to recognise different birdcalls, effectively measuring the activity of each bird species at specific times and locations.

The AI system uses audio that had been recorded and stored, chopping it into minute-long segments, and then converting the file into a spectrogram. A recurrent neural network ties together the chunks and produces a continual prediction of which of the three targeted birds is present across the minute-long segment.

The system was developed using Tensorflow, an open-source library for numerical computation for machine learning.

Birdcalls of Kākāriki, Hihi, and Tīeke, Victor Anton

Lessons learned

The audio was collected from forested areas, parks, and people’s backyards. So, the artificial intelligence algorithms have to deal with different sources of audio, not only birdcalls. At times the algorithms classify construction and traffic noises as birdsongs. Even doorbells have been misclassified as birdcalls.

Another challenge the researchers face is the variation of calls among birds of the same species. Birds have various calls and they use them for different purposes. For example, they will sing different calls when marking their territory than when they are looking for a partner. Training the system to identify specific call types is challenging because sometimes two or more birds are calling at the same time or the bird doesn’t sing the entire call.

Opportunities/Call To Action

The accuracy of the AI system improves as more audio is processed and more birdcalls are labelled. Researchers are looking at different options to provide the AI system with enough samples of birdcalls. For example, they are working on a citizen science project, where anyone would be able to help them identifying the birdcalls recorded in the audio.

The researchers are also working with other scientists and institutions, sharing the lessons they learnt and advancing towards faster and more accurate AI algorithms.

Next Steps

The researchers believe acoustic monitoring will be particularly useful to protect threatened species. Not only birds but other vocal animals, such as chimpanzees or whales, can be monitored using acoustic recorders. The combination of acoustic recorders and AI to identify the location and visiting frequency of threatened species could maximise current conservation efforts.

Third-party acknowledgment:

The research has been funded by Wellington City Council and Victoria University of Wellington.

Useful links:

Report abuse