AI Can Already Identify Mosquitoes by Sound. The Problem Is That Mosquitoes Are Not Consistent.

Mosquitoes kill more people than any other animal on Earth. The diseases they carry - malaria, dengue, chikungunya, Zika - account for millions of cases and hundreds of thousands of deaths every year. Knowing where mosquitoes are, and which species, could be transformative for public health: catch an outbreak of Aedes albopictus moving into a new region early enough, and vector control teams can respond before it establishes itself.

For years, researchers have been developing a surprisingly elegant tool for this kind of surveillance: listen for them. Every mosquito produces a distinctive sound when it flies - the buzz of its wingbeats. Females beat their wings slightly differently than males. Different species have different characteristic frequencies. And AI systems trained on recorded wingbeats can, in controlled conditions, achieve classification accuracy of up to 97 percent.

So why isn't passive acoustic mosquito monitoring standard practice everywhere? A study from Hungarian researchers, published in 2026, gets at the answer: the sounds are more variable in the wild than the training data suggests.

Recording hundreds of mosquitoes in Hungary

Scientists from the HUN-REN Centre for Ecological Research, ELTE University Budapest, and the University of Szeged captured and recorded mosquitoes from 10 of the most abundant species in Hungary. Their goal was systematic: quantify how much the acoustic signal varies between species, and how much individual variation exists within a single species, once environmental and biological factors are accounted for.

They assessed the impact of temperature, humidity, time of day, sex, age, and body size - measured by wing length - on the wingbeat frequencies produced by each species. Sound was generally consistent enough to discriminate between species. But the acoustic signal for a given species became even more reliable when those environmental and biological variables were controlled for. Which raises an uncomfortable question: what happens when they are not?

Temperature alone can shift a species' acoustic signature

Two variables stood out as particularly significant: sex and temperature.

Female mosquitoes had lower wingbeat frequencies than males across species, which makes biological sense - females are generally larger, and larger insects tend to beat their wings more slowly. That is a known complication for acoustic identification systems, which ideally need to tell species apart even when their most consistent fliers are not the ones doing the flying.

Temperature effects were more complex. Higher temperatures generally produced higher wingbeat frequencies - as temperatures rise, insect metabolism speeds up, muscles contract faster, and wings beat at higher rates. But the magnitude of this temperature effect differed between species. A temperate species and a subtropical species respond differently to the same temperature change. A species that primarily feeds on birds - whose blood runs cooler than mammalian blood - may have a different thermal response curve than one that targets humans.

"This species-specific difference in response to temperature suggests that we cannot apply a simple temperature correction rule for mosquito sounds, or at least that we cannot apply the same mathematical formula to all species," the researchers note. Any algorithm that corrects for temperature using a uniform formula will over-correct some species and under-correct others.

The training data problem

Current acoustic AI models for mosquito identification are trained on datasets that are often limited in species coverage and environmental variety. They may perform well on recordings made under controlled conditions but struggle when a sensor in the field encounters a mosquito that is hotter, younger, or smaller than any individual in the training set.

"Our data demonstrates that we cannot ignore intra-specific and intra-individual variability for AI-based acoustic classification," said Julie Augustin, the study's first author. "One solution for better integration of natural variance would be to adequately represent that environmental and biological variability in the training data. Unfortunately, such complete databases remain rare, especially for invertebrates, and building these extensive databases requires a lot of time and effort."

An alternative approach would be to build systems that take temperature and other environmental measurements as inputs alongside the acoustic signal - classifying not just on sound, but on sound in context. Some research groups are exploring this, but it requires detailed knowledge of how every species in the model responds to every relevant variable. That knowledge, for most mosquito species, does not yet exist.

Where this leaves field deployment

The practical implications are real but not fatal for the technology. Acoustic mosquito surveillance is still cheaper and less invasive than traditional trapping methods. The wingbeat-based approach can work in real time without laboratory analysis of physical specimens. And for high-priority species - those carrying the most dangerous pathogens, or those that are invasive and easily distinguished from background species - the accuracy may be sufficient even with natural variability.

What this study clarifies is that the gap between laboratory performance and field performance needs to be taken seriously before acoustic monitoring is deployed at scale for public health decision-making. Getting that gap wrong could mean false alarms or, worse, missed detections of emerging disease vectors. The mathematics of mosquito sound is manageable. The ecology of mosquito sound is harder.