Quantifying Biodiversity Through Sound Analysis

In an era where biodiversity is increasingly pressured by human activity and climate change, quantifying biodiversity is not just urgent—it’s imperative. Understanding the intricate web of life on our planet requires innovative methods.

Why Sound?

At RISE, we are approaching the issue through the modality of sound, something that has previously received less attention in AI research. Sound provides a rich, real-time stream of data from natural environments, offering insights into species presence, behaviors, and ecosystem health that are often unattainable through visual or other sensory data alone.

Novel AI Technology

We have introduced novel AI technology designed specifically to harness the power of environmental sounds for biodiversity monitoring. Our work builds on a number of building blocks, each pushing the research frontier in AI-based sound analysis:

Differentiable Log-Mel Spectrogram

We introduce the differentiable log-Mel spectrogram that dynamically adjusts spectrogram window lengths, optimized jointly with the neural network parameters.

Few-Shot Learning

We’ve incorporated few-shot learning, which minimizes the need for extensive labeled data, allowing our model to rapidly adjust to new, unique environmental sounds.

Active Learning with Adaptive Change Point Detection

Our combination of active learning and adaptive change point detection enhances data preparation efficiency, reducing the burden on experts.

Impact

Together, these contributions provide a strong platform for bioacoustic analysis and can be adapted to other application domains.

By bridging the gap between foundational AI research and practical applications in environmental science, our work not only advances technological capabilities but also provides vital tools for ecologists and conservationists. This dual contribution underscores the potential of AI to serve both the advancement of knowledge and the pressing needs of our planet.