notes

"Good morning everyone. We all know that biodiversity is changing rapidly across Europe. But before we can conserve biodiversity, we first have to measure it. Today, I'd like to show how machine listening is helping us scale biodiversity monitoring across multiple taxa, and why the biggest challenges are no longer the algorithms—but the data they learn from."

"Let's start with the problem."

"The biodiversity crisis is accelerating, but our ability to monitor it is not keeping pace. We simply don't have enough people, enough time, or enough resources to observe biodiversity at the scales conservation now requires."

"This gap between ecological change and monitoring capacity is becoming one of the major limitations for evidence-based conservation."

"Fortunately, ecosystems are already providing us with an enormous amount of information."

"Many organisms communicate acoustically. Birds sing, bats echolocate, frogs call, and grasshoppers stridulate. Passive acoustic monitoring allows us to record these soundscapes continuously over long periods and across large spatial scales."

"The challenge is no longer collecting recordings."

"The challenge is listening to them."

"This is exactly where machine listening comes in."

"Instead of manually analysing millions of recordings, AI can rapidly screen them, identify likely species, and even flag uncertain detections for expert validation."

"This transforms enormous audio archives into biodiversity information that can support monitoring, research and conservation decisions."

"But there is an important catch."

"Machine listening inherits the same limitations as machine learning in general."

Long-tailed taxa.

"First, training data are highly imbalanced. Birds dominate available recordings, while bats, amphibians and insects remain underrepresented."

Geographic coverage.

"Second, data are concentrated in a few regions, limiting how well models generalise across Europe."

Domain shift.

"Third, models trained on focal recordings often struggle when applied to passive acoustic monitoring."

"In other words: machine listening is ultimately limited by the data it learns from."

"FInally, one more challenge that we , most recordings only tell us which species are present—not when or where individual calls occur."

"Most community recordings come with a single label for the entire clip."

"The model knows that Brown long-eared bat occurs somewhere in this recording—but not where."

"That makes learning much more difficult."

"Strong labels go much further."

"Here every vocalisation is annotated with its timing, frequency range, vocalisation type and even background species."

"These annotations are expensive to produce—but they provide exactly the information modern machine learning needs."

"One of our goals is making these annotations the standard."

"Our work addresses several of these limitations."

Then briefly:

• multi-taxa datasets

• birds

• bats

• amphibians

• orthopterans

"Rather than focusing on birds alone, we're building open datasets and recognition models across four major European taxonomic groups."

Then

Xeno-Canto.

"Through our collaboration with Xeno-Canto we're also helping transform weak community labels into the strong annotations needed for the next generation of models."

"Finally, I'd like to show one application that we're particularly excited about."

"Instead of deploying stationary recorders, we can simply walk a transect while recording audio."

"Machine listening automatically links species detections to GPS positions, producing georeferenced biodiversity observations."

Advantages.

"This allows rapid multi-taxa surveys without requiring specialist taxonomic expertise."

Limitations.

"It won't replace standard monitoring schemes—it doesn't estimate breeding abundance and it depends on vocal activity—but it offers a promising complement for rapid biodiversity assessments."

I'd like to leave you with one final thought. Machine listening is not about replacing ecologists or field experts. It's about making their expertise scalable.

And I think that's exactly what we'll need if we want biodiversity monitoring to keep pace with the biodiversity crisis. Thank you."