A new paper from the Responsive Environments group puts real-time bee identification on a low-power chip inside the BuzzCam field recorder.

In the temperate rainforests around Nahuel Huapi National Park in Argentine Patagonia, the world’s largest bumblebee is becoming harder to find. Bombus dahlbomii, known locally as the moscardón, was once common across southern Chile and Argentina. After the European bumblebee Bombus terrestris was imported for crop pollination in the late 1990s and spread through the region, the native species declined so sharply that it is now listed as Endangered on the IUCN Red List.

Tracking that decline means knowing where the bees still are. Traditionally, that has required expert observers working in remote terrain. Passive acoustic recorders can listen continuously, but they produce large audio archives that must be stored, retrieved, and analyzed long after the moment has passed. The power and storage this demands limits how long, and how widely, they can be deployed.

Recognition on the sensor itself

In a paper published this week in Scientific Reports, researchers from the Media Lab’s Responsive Environments group and collaborators in Argentina describe a system that runs the analysis in the forest, on the sensor itself, as the buzz happens. The team trained a compact convolutional neural network (158,144 parameters, quantized to 8-bit precision) to tell the flight buzz of B. dahlbomii apart from that of invasive B. terrestris and from background sound. The model runs on a low-power microcontroller with a built-in neural network accelerator, integrated into the group’s BuzzCam field recorder.

On held-out test data, the model distinguished the three classes with 86.1 percent accuracy. It classifies each one-second audio segment in 10.4 milliseconds and each classification consumes just 794 microjoules.

Because recognition happens on the device, a sensor no longer needs connectivity, cloud computing, or a researcher’s return visit to say something useful. It can log species presence in real time, record only when a bee is actually present, and make long-term, many-site monitoring networks practical in places where hauling data or people is the hard part.

“A buzz carries a surprising amount of identity. These two species sound different if you know how to listen. What excites me is that the sensor doesn’t just record sound anymore; it understands it in the moment, on an energy budget small enough that you can leave it in the forest and walk away,” says first author Patrick Chwalek, a research affiliate in the Responsive Environments group who began the work as a PhD student and continued it as a postdoc.

Built on an open dataset

The model was trained on data the team collected and published openly last year in Scientific Data. In March 2024, working with Argentine and Chilean ecologists and supported by the National Geographic Society, the group deployed nine BuzzCam recorders at twelve sites near Puerto Blest, on the shore of Nahuel Huapi Lake. Over six days the devices captured roughly 250 hours of stereo audio along with temperature, humidity, and air-quality readings, while field observers tagged bee visits in real time through a custom iOS app. After crowdsourced validation, the effort yielded more than 21,000 labeled one-second clips, one of the richest acoustic records of these species assembled to date. The dataset is freely available to other researchers.

The team

Chwalek’s co-authors on the new paper are Marie Kuronaga (Media Lab and Kioxia Corporation), Marco Giordano (ETH Zurich), Aidan Bradshaw and Isamar Zhu (Media Lab), Professor Joseph A. Paradiso, who directs the Responsive Environments group, and Marina Arbetman of INIBIOMA (Universidad Nacional del Comahue and CONICET) in Bariloche, Argentina. The underlying dataset was built with additional collaborators at the Universidad Metropolitana de Ciencias de la Educación in Chile, the Jožef Stefan Institute in Slovenia, and the University of Missouri. BuzzCam was named an honoree in Fast Company’s 2025 World Changing Ideas Awards.

This research was supported by the National Geographic Society and its Exploration Technology Lab, a Google Research Award, and Kioxia Corporation.

Related content

• Paper: On-device deep learning for real-time acoustic monitoring of endangered Bombus dahlbomii and invasive congeners, Scientific Reports (2026)
• Dataset paper: High-Res Acoustic and Environmental Data to Monitor Bombus dahlbomii Amid Invasive Species, Habitat Loss, Scientific Data 12, 548 (2025)
• Project: BuzzCam
• Gallery: Bumblebee Data Collection in Puerto Blest, Argentina (March 2024)
• Member collaboration story: Responsive Environments Group and Kioxia Collaborate on Wildlife Sensing Technology (Nov. 2024)
• Post: Fast Company’s 2025 World Changing Ideas Awards: Media Lab Honorees