Guide · AI & data
AI for Forest Elephants: Listening for Rumbles and Gunshots
By the FruitPunch research deskUpdated November 15, 20268 min read
In short
African forest elephants were listed as Critically Endangered by IUCN in 2021 after their numbers fell by more than 86% over 31 years. The Cornell Elephant Listening Project runs a grid of 50 recorders over 1,250 km² of northern Republic of Congo, and machine learning detectors cut the audio needing human review by 98% in one Stanford and Cornell study. Gunshot detectors are now small enough to run on the recorder itself, with one 2026 model at about 935k parameters and 964KB.
This address once hosted FruitPunch AI, an Eindhoven AI for good community. FruitPunch is a new, independent publication and is not that organization. This guide was written from public sources; the history page explains the change.
AI helps protect African forest elephants by listening for them. Forest elephants live in dense Central African rainforest where they are hard to see and count, and the IUCN reported in 2021 that their numbers "fell by more than 86% over a period of 31 years", moving the species to Critically Endangered. Researchers now leave grids of sound recorders in the forest for months and use detection software, increasingly deep learning, to pick elephant rumbles and gunshots out of hundreds of thousands of hours of audio. One study found that a neural network reduced the audio requiring human inspection by 98%.
Key facts
- 86% decline: forest elephant numbers fell by more than 86% over 31 years, and the species is listed as Critically Endangered on the IUCN Red List.
- Main cause: a 2025 study in Philosophical Transactions B says the decline between 1990 and 2021 was mainly because of ivory hunting.
- 50 recorders over 1,250 km²: the acoustic grid in and around Nouabalé-Ndoki National Park in the Republic of Congo, described by WCS Congo as one unit every 5 km².
- More than a million hours of sound collected since 2017, according to National Geographic, now shared on AWS as an open dataset.
- 0.5% precision: the template gunshot detector used in the forest in 2021, which student teams then beat with object detection models reaching 95% precision.
Why are African forest elephants critically endangered?
The African forest elephant (Loxodonta cyclotis) was assessed as a separate species from the savanna elephant in the 2021 IUCN Red List update. The IUCN announcement gave two figures side by side: forest elephants fell by more than 86% over 31 years, while savanna elephants fell by at least 60% over the last 50 years. Both species declined sharply after 2008 because of a significant increase in poaching, which peaked in 2011, and the IUCN names conversion of habitat to agriculture and other land uses as another significant threat.
Poaching also changes how surviving elephants behave. A 2025 Elephant Listening Project study in and around Nouabalé-Ndoki found that elephant occupancy drops from 0.54 to 0.52 after a gunfire event, a change sustained over 8 days, and that more gunshots led to a significant rise in the share of night-time vocal activity. A 2024 PLOS ONE paper from the same grid reports that forest elephants in relatively undisturbed forest are active nearly equally day and night, but become more nocturnal when exposed to a perceived risk such as poaching.
Why use sound instead of cameras or surveys?
Forest elephants are hard to observe in dense rainforest, and traditional surveys are slow. A 2017 paper in the Journal of the Acoustical Society of America notes that forest elephants occupy large ranges and often use far-reaching vocal signals, and that the methods used to monitor Central African populations are expensive and time-intensive. Sound travels through vegetation that blocks a camera's view, and a recorder works day and night.
The review by Wrege, Rowland, Keen and Shiu in Methods in Ecology and Evolution argues that passive acoustic monitoring "is a viable and cost‐effective tool for conservation". The same paper names the bottleneck that AI now addresses: processing the dense raw data efficiently with standard computer hardware and speeding up the development of detection algorithms.
How does the Elephant Listening Project grid work?
The Elephant Listening Project, now part of Cornell's K. Lisa Yang Center for Conservation Bioacoustics, works with the Wildlife Conservation Society (WCS) Congo Program in northern Republic of Congo. WCS describes the network as 50 acoustic sensors, randomly distributed at a density of one unit every 5 km², covering part of Nouabalé-Ndoki National Park, the Kabo forest concession and a safari zone.
Recorders run continuously and are collected in rounds. The Imperial College team that worked on the project described the deployment as one recorder per 25 square kilometers, with collection taking 21 days every 4 months at the time, with audio stored on SD cards. The 22nd deployment ran from August 2025 to April 2026: 46 acoustic units logged more than 36,156 hours and detected no gunshots, and WCS uses the results to direct eco-guard patrols.
The data has also shown when protection lapses. During a three-month ranger training period in early 2018, when patrols were reduced, recorded gunshots spiked, and the audio showed poaching deeper in the park than expected.
How do AI rumble detectors find elephant calls?
Most detectors turn audio into a spectrogram, an image of frequency over time, and then classify short windows as "rumble" or "not rumble". Forest elephant rumbles are low frequency and varied: a study of 686 rumbles from four sites across Central Africa found they could be classified into five to eight types, with highly intergraded structure.
The published detectors show steady progress:
| Study | Method | Data | Reported result |
|---|---|---|---|
| Keen et al., JASA 2017 | Automated rumble detector | Nearly 4,000 h from five forest clearings in Central Africa | 83.2% true positive rate at 5.5% false positive rate (about 20 false positives per hour) |
| Bjorck et al., AAAI 2019 | Deep learning detector plus learned audio compression | Crowdsourced labels on thousands of hours | Improved call detection and segmentation over earlier methods |
| Gomes-Selman et al., ESA 2020 | Neural network | Nouabalé-Ndoki recordings | 98% less audio needing human inspection |
| Geldenhuys and Niesler, 2024 preprint | Audio spectrogram transformer | Continuously recorded elephant audio | Average precision of 0.962 for framewise call detection |
The Bjorck et al. paper, by Cornell computer scientists working with the Elephant Listening Project, also tackled a practical issue: common lossy audio codecs are tuned for human hearing and handle low-frequency elephant calls poorly, so the authors proposed a trainable compression method. Smaller files matter when audio has to be sent from the forest over limited links.
Commercial partners have helped with scale. NVIDIA's developer blog reports that Conservation Metrics, a Santa Cruz startup, used deep learning to detect elephant calls and gunshots from Nouabalé-Ndoki, analyzing over seven terabytes of recordings every three months for that project alone.
How do gunshot detectors work, and why are they hard?
A gunshot is short and loud, but rainforests are full of sounds that look similar in a spectrogram. The older template detector used by the project had only 0.5% precision, so almost every alert was a false alarm. A 2021 Imperial College London student team, working through Microsoft's Project 15, tested several models on more than 400 hours of forest clips. Their reported gunshot precision was:
| Model | Precision |
|---|---|
| Template detector in use | 0.5% |
| Azure Custom Vision | 60% |
| YOLOv4 | 85% |
| Faster R-CNN | 95% |
Speed was the trade-off: Faster R-CNN took about 100 minutes per 24-hour file, against 44 minutes for YOLOv4, according to the same project page.
Outside Central Africa, a widely used benchmark comes from Belize. Katsis et al. in Ecological Indicators (2022) trained a spectrogram CNN on 597 gunshots and 28,195 background sounds. It reached 0.95 recall and 0.85 precision on validation data, but precision fell to 0.57 on over 160,000 files from a different site. That drop is the central problem for any gunshot model moved to a new forest.
Can the AI run on the recorder itself?
Yes, and this is where the work is heading. Running a detector on the device means it can save only relevant clips, use less power and storage, and in future raise alerts without waiting months for SD cards to be collected.
- AudioMoth. The AudioMoth is a low-cost, open-source logger powered by 3 AA batteries that records at 8,000 to 384,000 samples per second. Its developers published on-device detection algorithms in Sensors (2019), including one to collect evidence of poaching, and benchmarked a hidden Markov model against a CNN.
- Lightweight gunshot CNN. In a 2026 paper in Biodiversity Information Science and Standards, Naveen Dhar describes a model with approximately 935k parameters that shrinks to 964KB after TensorFlow Lite conversion, with 0.96 average precision on the Belize validation data.
- Networked filtering. Dhar's December 2025 presentation with the K. Lisa Yang Center and the Elephant Listening Project, reported by the Acoustical Society of America, proposes running the model on each sensor's microprocessor and a central hub that checks whether other sensors heard the same sound before treating it as a gunshot. Field testing had not yet happened at the time.
Projects, datasets and tools you can use
- Sounds of Central African landscapes. The Elephant Listening Project's archival soundscapes from Nouabalé-Ndoki on the AWS Registry of Open Data. The license allows scientific study and development of detection algorithms, with attribution to the "Cornell Elephant Listening Project".
- Dryad dataset from the 2024 PLOS ONE study. Hourly rumble detections and gunshot events from 50 recorders, 15 January 2017 to 27 April 2021, including 980,691 rows of rumble observations and 86 gunshot event rows.
- Project 15 gunshot detection documentation. The Imperial College team's design notes and code resources.
- OpenSoundscape. An open-source Python package with tutorials for training CNNs on spectrograms and for building classifiers on BirdNET or Perch embeddings.
- AudioMoth. Open hardware recorder used in many low-cost deployments.
- Belize gunshot benchmark. The Katsis et al. dataset, used by later lightweight models as a pre-partitioned set of 35,980 four-second recordings.
What do the results show so far?
The acoustic record now does more than count calls. It has shown that elephant presence drops for 8 days after hunting incidents, that logging is perceived as a risk on moderate time and small spatial scales but with little effect on elephant density, and that a full 2025 to 2026 deployment can pass without a single gunshot detected. The limits are also clear: detectors trained at one site lose precision at another, and most data still reaches analysts months after it is recorded.
Frequently asked questions
How many African forest elephants have been lost?
The IUCN says forest elephant numbers fell by more than 86% over a period of 31 years, which led to the 2021 Critically Endangered listing. Poaching for ivory and habitat conversion are the main threats.
What is the Elephant Listening Project?
It is a research program at Cornell's K. Lisa Yang Center for Conservation Bioacoustics, established in 1999, that uses sound to study and protect forest elephants. With WCS Congo it runs a 50 recorder grid over 1,250 km² around Nouabalé-Ndoki National Park.
How accurate are elephant rumble detectors?
A 2017 detector reached an 83.2% true positive rate at a 5.5% false positive rate on nearly 4,000 hours of Central African recordings. A 2024 transformer model reports average precision of 0.962 for framewise call detection, though on different data, so the numbers are not directly comparable.
Can AI detect poachers in real time?
Not yet at scale in Central Africa. Current grids store audio for months before analysis. Lightweight models of about 964KB can run on sensors, and a networked design that cross-checks detections between sensors was still under development in December 2025.
Where can I get forest elephant audio to train a model?
The Elephant Listening Project publishes Nouabalé-Ndoki soundscapes on the AWS Registry of Open Data for scientific use and algorithm development, and the 2024 PLOS ONE study shares its rumble and gunshot detections on Dryad.
Spotted an error in this guide? Email editorial@fruitpunch.ai