Support Provided ByLearn More

‘Seeing’ Nature’s Voice

Can artificial intelligence help wildlife biologists find the last of a rare and treasured species of frog?

ByAlex RileyNOVA NextNOVA Next
‘Seeing’ Nature’s Voice

Jan Paul Zegarra points toward a mound of dried dirt that is home of thousands of fire ants, each one with a dark ruby tail containing a stinger and burning venom. “Welcome to Carite Forest,” he says. “You don’t want to step on that.”

After driving up the steep and potholed roads that lace Puerto Rico’s central mountains, Zegarra had parked his white SUV just over the nest, as if he was setting the stage for the first lesson of what to avoid in a rainforest. The second: a tree called the manchineel that produces sweet-tasting, but highly toxic, apple-like fruits and coats its every surface—bark and leaves—in milky sap that that burns the skin and can temporarily blind the eye. In Spanish, it is known as manzanilla de la muerte or “little apple of death.”

Jan Paul Zegarra and Recorder
Biologist Jan Paul Zegarra mounts a recorder to a tree in a quest to find evidence that a rare frog still lives.

Briefed on what to avoid, and wishing that I had worn a long-sleeved shirt and perhaps a pair of goggles, we start discussing the reason we are both here: a tiny frog. Known as the golden coquí, its color actually ranges from olive to gold. It’s small enough to just about fit on a quarter, and it’s native to this small island—perhaps even this forest. First discovered in 1974 and not seen or heard after 1983, it is at best critically endangered and at worst extinct. Surveys in the 1990s came back negative, but the areas that were sampled weren’t far from the well-trodden paths. With a handheld GPS device as a guide, Zegarra and I were going much deeper, to places untouched for decades, where large bromeliads—the plants in which the golden coquí was once found—might thrive.

At seven promising sites, Zegarra will strap a sound recorder to a tree and move on, extending our sense of hearing in both space and time while allowing the forest’s natural rhythm to continue undisturbed. For a week, these devices will switch on for one minute every ten minutes, recording the sounds from within a few hundred yards of the microphone—rain, insects, birds, and, importantly, frogs. In total, Zegarra aims to strap over 60 recorders throughout the Carite Forest, amassing around 900 hours of unfiltered audio. That’s over 37 days’ worth of data, which would take months, if not longer, to process. So rather than listening to it all, patiently waiting for a frog to sing, Zegarra and the U.S. Fish and Wildlife Service for whom he works have hired a helping hand: artificial intelligence.

For the last decade, Puerto Rico-based company Sieve Analytics has been training computers to identify the distinctive calls of different species across the globe. From gold mines in Colombia to forests in Hawaii, and from the top of mountains in Puerto Rico to beneath the waves of the Pacific Ocean, their technology has been tracking the presence and absence of certain species and, by proxy, the health of those ecosystems. Over time, the algorithms have become more precise, better able to automatically extract species from background noise. “If the frog is out there and it is singing,” says Marconi Campos-Cerqueira, a Sieve Analytics employee, “we’ll find it.”

Encoding a Superpower

Campos-Cerqueira is short, only five feet, five inches tall. He has a fluffy black beard, and, he says, a superpower. Like many good comic book characters, his developed in the face of disability. Born color blind and developing acute shortsightedness in his teens, he doesn’t sense the world quite like most people. Often, he uses his ears more than his eyes.

This became very apparent in 2012 when he was working in the Amazon rainforest in Brazil, not far from where he was born, trying to identify the hundreds of bird species that surrounded him. “When I first arrived in the Amazon, I was so excited because I thought, ‘Wow, I will definitely see a lot of birds and jaguars and tapirs,’” he says. “But it was so impossible to see them.”

“If the frog is out there and it is singing, we’ll find it.”

And not because of his poor vision. A rainforest is a web of foliage—light doesn’t travel far. But when he stood under the canopy, Campos-Cerqueira could close his eyes and take in his surroundings with his ears. In a way, it was easier. Not only could he determine which species were around him, but where they were—both in direction and distance. There are over 400 species of birds that inhabit the Amazon, a maze of squeaks, squawks, clicks, hoots, and whistles. But his ears found a way through the cacophony, pinpointing each call to its creator.

As useful as this was, his technique still had its shortcomings. He couldn’t spend every day in the rainforest in a Zen-like state plugged into nature’s soundtrack. Plus, he wanted to compare different sites, exploring how human development altered the soundscape. But, he says, “I was poor.” So he made his own recording devices from old mp3 players and battery packs. Across 150 sites in the Amazon, he could sample species all day long and then listen to them on his computer.

After four months of recording, he had five terabytes of data to pore over—far too much for one person to filter through. He reached out for help from students at the National Institute of Amazonian Research, but quickly realized that few people could accurately identify all the bird species in the Amazon. He had all the data he could dream of, but no easy way to assess it.

That’s when he met Mitchell Aide. A professor of land use and ecology at Puerto Rico University, Aide was in Brazil in late 2012 to give a talk about the automation of acoustical surveys. Since 2008, he had been training computers to take big data and only extract the parts that were useful to a project’s aims. From a burden of noise, he is provided with species and their sounds. He called this side project Sieve Analytics.

Mitchell Aide and his colleague, Orlando
Mitchell Aide and Orlando Acevedo-Charry study sound data in search of a particular sound print.

“I just loved what he was doing,” Campos-Cerqueira says. “At that time, it was a huge novelty to be able to automatically identify sounds.” After they met in Brazil, Campos-Cerqueira boarded the plane to Puerto Rico, where for a month, he learned how train a computer to identify his bird calls.

He had arrived at a perfect time. Although acoustic ecosystem surveys were nothing new, the idea of automation was a novel addition that was maturing quickly. The original software Aide’s programmers developed in 2007, known as the “Automated, Remote, Biodiversity, Monitoring Network,” or ARBIMON, ran off hard drives stashed in his office, sometimes leading to the headache of irretrievable data loss. When I asked how much data was lost, the programmer replied that it was so bad, he had tried to forget the whole event. The user interface was not very user friendly, either. Yet interest in Aide’s product grew, and soon he decided it was time for an update. In 2014, three programmers working full-time for six months developed ARBIMON II, a piece of software that lives in the Amazon Cloud, a far-off suite of harddrives that guarantee a reassuring 0.0001% chance of data loss. And with a flashy new interface and mobile app, it is easy to use.

Work in ARBIMON II starts with a spectrogram, the visual representation of a sound. As we sat in Aide’s office one sweaty afternoon, he showed me a recording from the wetlands of Puerto Rico. I could see the distinctive, and repetitive call of one species of frog, known as the wetland coquí: two diagonal stripes at five kilohertz that was akin to the claw mark of a two-toed animal. Aide drew a box around the soundprint—not unlike highlighting a person’s face in a photo—and committed the shape to ARBIMON II’s memory. After repeating this process with other soundprints from different individuals of the same species, the software creates a composite soundprint that’s representative of the species. With this composite, the software is ready to start scanning through messy jungles of data.

ARBIMON II uses what’s known as machine learning, a basic form of artificial intelligence. In recent years, AI has made inroads in conservation, from robots that can hunt and kill invasive species to drones that can reseed a denuded ecosystem with native plants. They share a common goal: extending the reach of humans into ecosystems without a human actually being there.

Support Provided ByLearn More

Seeking the Golden Coquí

Back under the canopy of Carite Forest, Zegarra and I are making our way—slowly—through the tangle of palm tree trunks and low-hanging ferns, dried leaves the size of bin lids crunch under our feet like giant poppadums. In this place, a 10-yard walk can take over an hour as we tread carefully, avoid the most conspicuous spiders, and capture photographs as we go. We’re not actually looking for a golden frog, but the treehouse-like bromeliads the frogs are thought to dwell in. “The critical factor determining the presence of populations,” wrote George Drewry and Kirkland Jones, the two scientists who discovered the frog, in 1974, “appears to be dense bromeliad growth, which in turn seems to be dependent on heavy dew produced by [updrafts of air] striking the mountain range.”

Such bromeliads have become increasingly rare because the rainforest is becoming progressively drier as the flow of moist air from below has weakened over the years.

Bromeliad Paradise
An opening in the understory reveals a startling number of bromeliads.

The bromeliads represent a best guess as to where the golden coquís may be. Yes, they were originally found within them, but that doesn’t mean they are always in the plants. Perhaps they prefer streams, ponds, or another type of water-pooling plant, Zegarra tells me. In fact, what is known about the golden coquí is actually very little. It has only been found in one forest in Puerto Rico, preferring a band of altitude between 2,120 and 2,575 feet. Females give birth to live young, retaining their eggs until they hatch into little froglets with no tail and fully-formed legs. And, importantly, males only sing between midnight and dawn, chirping a soft triplet of whistles that could easily be confused for a bird to the untrained ear.

Even with recorders that extend our senses throughout the day, searching for the frog is like finding a needle in a haystack—if the haystack was the size of a mountain that may or may not contain a needle. So why are we looking? For some, including Zegarra, the frog represents something much larger than its population size. Like him, the species was born on Puerto Rico, raised on its water and food, and never left. Plus, the frog is one of only 14 species of coquí that are found on Puerto Rico—and it is the golden one! For herpetologists of the 1970s, it was the frog’s odd—but not unique—reproductive biology that attracted them to the forests, taking hundreds of individuals back to their labs on the island and to the United States. Seven years after it was first discovered inside bromeliads, the species disappeared from all human senses.

At one site we strike gold, though not of the coquí variety: a pool of water surrounded by an amphitheater of trees covered in bromeliads. “It’s like bromeliad heaven in here,” Zegarra says, mouth agape as he gazes above. To a tree trunk, he straps a recorder—an external microphone hooked up to a smartphone with the ARBIMON app installed, all kept dry inside a hard plastic case. He repeats this at six other promising sites throughout the day. At one, we come across Coccoloba rugosa , with its long thin trunk, large lily pad-like leaves, and, at this time of year, a scarlet inflorescence of flowers erupting from the top. To Zegarra, it is a floral flag of a healthy habitat. C. rugosa doesn’t like change, only growing in areas that haven’t been disturbed for decades, Zegarra tells me. Near one of the recording sites, one of the plants has been bent down to our waist height, as if bowing to its adoring fans. Large orange and black butterflies land on its tiny flowers, unfurling their tongues while Zegarra focuses his macro lens. In all his years on Puerto Rico, he had never seen so many in bloom before.

Beyond the Golden Coquí

In 2013, as part of his PhD project with Aide at the University of Puerto Rico, Campos-Cerqueira demonstrated that another rare and endemic species of Puerto Rico known as the Elfin-woods warbler was actually more common than once thought. His recorders—placed up the mountain from bottom to top—found that it was more likely to be found in the Palo Colorado forest, the ecosystem that lies at lower elevations than the Elfin woods. Previous surveys may have missed this, he says, since the bird doesn’t sing at the dawn chorus with the majority of other species, but takes its chance at midday when the sun is brightest and the forest is quieter.

“We used to think that birds—all birds—were vocally active during the dawn chorus,” Campos-Cerqueira says. “But that’s not true.” A species can appear to be rare, he says, “because you were going everyday—for all these years—during the dawn chorus expecting it to be there, but it’s not.”

With the recorders out in the field 24/7, the whole soundscape is saved for future analysis. As Aide laconically puts it, “The recorders record everything.”

“We want to be 99.999% sure that we have not missed the golden coquí in the recordings, especially given that this maybe the last time people search for the species.”

To train a computer to identify the call of the golden coquí, Campos- Cerqueira had to be inventive. There was only one sound recording for the species, and a single individual isn’t the perfect representative for a whole species, like using one face to build face-recognition software. To overcome this problem, he melded this one audio file with background noise from Carite Forest, increasing and decreasing its volume to put his computer models to the test. After this training regimen, his computers then scoured through the datafiles from Carite Forest looking for anything like the golden coquí.

Within a couple of hours he had had his results: nine possible hits. But, on closer inspection they were false positives, not the golden coquí. Unfazed, Campos-Cerqueira tried again with a slightly more flexible model, one that will pick up anything that is remotely similar to the golden coquí call and can then be then checked by a human. “We want to be 99.999% sure that we have not missed the [golden] coquí in the recordings, especially given that this maybe the last time people search for the species,” Aide told me in an email.

Whatever the outcome of the golden coquí project, Aide hopes his technology will provide a service to saving species in the future, when a reduction in a species’ call sounds an alarm. He likens such global network of biodiversity monitors to the current system of weather stations, which can be used both to analyze the present and to forecast the future. As species adapt and move in response to climate change, for instance, his recorders would detect the loss or shifting population. With long-term data, annual and even decadal rhythms can be decoded into simple numbers.

Despite their omnipresence, there are still doubts that ARBIMON II is better than a human in the field. A 2006 study by Miguel Acevedo, one of Aide’s students who helped found ARBIMON, concluded that AI was better, identifying two more species than human surveys. “They could record things that we can’t hear, because we weren’t there,” says Acevedo, who now studies infectious diseases in lizards at the University of Puerto Rico. A study in 2013, however, found the opposite: humans were slightly better at recording the number of species, relying on both sight and sound.

Whatever the case may be, these direct comparisons miss the point, Aide tells me as we paddle board on one of San Juan’s lagoons early one morning looking for manatees, another one of Puerto Rican’s rare residents. To him, they complement one another: some jobs require an expert, he says, but others are cheaper and more effective if outsourced to a recorder and ARBIMON II.

Conservationists are certainly far more experienced than computers, but they also have limited time and money. Leaving a few recorders in the forest is much cheaper than paying for a suite of experts—a herpetologist for frogs, an entomologist for insects, an ornithologist for birds.

The field has been opening up to the idea. At a recent conservation conference, one attendee said that ARBIMON II is “breaking new ground in biodiversity research.” The website currently has over 600 users, past and present, many with their data open to anyone else that might be interested in looking at their ecosystem’s acoustic health. And, perhaps there is no greater dedication and trust in ARBIMON II than allowing it to decide whether a species is alive or extinct, whether there is hope for the future or a full stop in one of life’s replicating codes.

Slim Odds

In nearly 40 days of recordings from over 60 sites—some of which seemed idyllic for a little frog—the golden coquí didn’t call. There was just background noise where there was once a song, a soft triplet whistle.

Campos-Cerqueira admits that the golden coquí’s odd at survival were slim from the start. The chytrid fungus that is drying up the permeable skins of amphibians throughout the tropics, diluting their populations, had settled in Puerto Rico as early as 1976, just two years after the golden coquí was discovered. Plus, the island sits in the path of two highly destructive forces of nature: hurricanes from the Atlantic Ocean and humans from all over the world. Combined with the frog’s inability to spawn large numbers of offspring, while being collected throughout the 1970s largely for this trait, it’s not that surprising that the recorders didn’t pick up its song.

While driving into the rainforest on dirt tracks, Zegarra, the field biologist from the USFWS, told me about a Puerto Rican farmer who found golden coquís on his land. Their presence made his property very popular with the local people. Every day, more and more people would arrive, and fewer and fewer frogs would remain in the bromeliads. As this continued over weeks, the landowner decided to make a stop to his unintentional attraction. He set fire to his lands, burning the rainforest from parasol-like palm tree to low-hanging fern. The visitors stopped bothering him. The frogs died. Regardless of whether this happened or not, it is certainly one of the more powerful stories that will outlive the golden coquí. Through word of mouth, the frog lives on as just another species that was at the centre of forces much larger than itself, each one coalescing like flames through a forest.

Receive emails about upcoming NOVA programs and related content, as well as featured reporting about current events through a science lens.


On September 6, Hurricane Irma hit Puerto Rico as a Category 5 storm, pounding the island with winds of 185 mph, and then on September 20, Hurricane Maria made landfall, battering the already beaten island. The two storms downed powerlines, tore homes to pieces, and denuded entire hillsides. The previously quiet 2017 hurricane season had become an humanitarian and ecological disaster.

Coast Guard crews deliver supplies to remote location in Puerto Rico
Entire hillsides in Puerto Rico were stripped of their greenery in the wake of two powerful hurricanes.

Aide and his colleagues hope to use their acoustic surveys to study how the islands forests are recovering. “By this time next year, we should have a better idea of the impact of the hurricane,” he tells me.

It’s likely that many populations will bounce back, but for one little frog, hopes have dimmed further. “I do not doubt that vulnerable species with small and few populations will be impacted by hurricane Maria,” Aide says. Basically, if Aide, Zegarra, and their colleagues did miss a small population of golden frogs during their wide-ranging surveys, the chances that they are still alive now is even slimmer.

Photo credits: Alex Riley, George Dewey/USFWS, Eric D. Woodall/U.S. Coast Guard

Funding for NOVA Next is provided by the Eleanor and Howard Morgan Family Foundation.

National corporate funding for NOVA is provided by Draper. Major funding for NOVA is provided by the David H. Koch Fund for Science, the Corporation for Public Broadcasting, and PBS viewers. Additional funding is provided by the NOVA Science Trust.