A Neural Network Identifies Termites Inside Your Wall by Listening to Them Eat. Your Inspector Uses a Screwdriver.
A hybrid CNN-LSTM acoustic model detects active termite infestations through wood with 94.5% accuracy and a 95.8% recall rate. The Termatrac T3i, the best commercial sensor on the market, still requires a human operator to interpret raw waveforms and flags false positives from passing cars. The standard WDI inspection for a home sale costs $200 and catches roughly a tenth of active damage. The gap between what works in a lab and what shows up at your closing table costs American homeowners $6.8 billion a year.
Somewhere in a lab, a machine learning model is listening to termites eat.
Not metaphorically. A piezoelectric sensor, pressed against a block of pine, picks up the acoustic signature of mandibles shearing cellulose fibers at frequencies between 1 and 100 kHz. Raw audio gets sliced into Mel-Frequency Cepstral Coefficients, the same feature extraction technique that powers speech recognition on your phone. A convolutional neural network learns the spatial patterns in those coefficients. A Long Short-Term Memory layer learns the temporal patterns, because termites chew in bursts, not continuous streams. Published on arXiv in July 2025, the hybrid CNN-LSTM architecture classifies wood samples as infested or clean with 94.5% accuracy, 93.2% precision, and 95.8% recall.
That recall number matters. It means the model misses roughly one in twenty active infestations. A human inspector, armed with a flashlight and a flat-head screwdriver, misses closer to nine in ten.
What a $200 Inspection Actually Detects
When you buy a house in Florida, Texas, Georgia, the Carolinas, or any of the thirty-plus states where lenders require a Wood-Destroying Insect report before closing, the inspection follows a pattern that has not fundamentally changed since the 1970s. A licensed pest control operator walks the exterior perimeter, crawls through the crawlspace if one exists, checks the attic framing, probes exposed wood with a screwdriver, and looks for the visual evidence of infestation: mud tubes built by subterranean termites, frass pellets kicked out by drywood colonies, hollow-sounding wood, sagging floors, bubbling paint.
Termites are good at hiding, which is the problem. Subterranean colonies can number in the millions and eat continuously for years before producing a single visible sign above the soil line. Drywood termites push frass through kick-out holes the size of a pinhead that an inspector has to spot in a dimly lit attic from four feet away while crouched between HVAC ducts. By the time damage is visible to the naked eye, structural loss is already significant.
The National Pest Management Association puts annual US termite damage at $6.8 billion. Dr. Nan-Yao Su at the University of Florida, who has studied termite economics since the 1980s, calculated a broader figure of $20 billion when you include control costs, indirect property losses, and the repair-to-treatment multiplier he estimated at 5:1. A PCT Online industry survey quotes an unnamed structural engineer: "You probably only see about 10 percent of termite damage, and most homeowners only fix what they see."
Standard homeowner's insurance excludes termite damage entirely. VA and FHA loans require a clean WDI report to close, but neither specifies what detection technology the inspector must use. A screwdriver meets the standard.
The Machines That Exist and Nobody Uses
Acoustic termite detection is not a new idea. Researchers at CSIRO in Canberra published studies on it in the early 2000s. At the USDA Forest Products Laboratory in Madison, Wisconsin, an active research program explores bio-acoustic termite behavior, including whether playback of ant footstep vibrations can actually deter colonies from entering structural timber. Underneath all of it is mature science: termites generate detectable sound when they chew, when they head-bang to signal alarm, and when they walk. Piezoelectric sensors, accelerometers, and MEMS microphones can all pick it up.
Two meaningful products exist on the commercial market.
First is the Termatrac T3i, an Australian device introduced in 2010 as a successor to the original T1r from 1999. It uses low-energy microwave radar, not acoustic sensing, to detect motion through wood and other building materials. It also includes a moisture meter and thermal sensor. It works, but with significant limitations that researchers at UC Davis and the University of California Agriculture and Natural Resources have documented in field trials.
Termatrac picks up motion from cars passing outside, people walking in adjacent rooms, pets, water flowing through pipes, ants, and other insects. Its signal attenuates with depth, producing false negatives on infestations deep inside a wall or below a concrete slab. Interpreting its raw line-graph output requires training and experience that most pest control operators do not have. Siavash Taravati's 2018 field evaluation, conducted across infested structures in Southern California, found that experienced operators could achieve reasonable accuracy with the device, but that inexperienced users produced results barely better than chance.
Second is the WiSPr, a wireless acoustic sensor network developed by Associate Professor Adam Osseiran's team at Edith Cowan University in Australia. Its concept was compelling: place twenty fingernail-sized acoustic sensors around a home's perimeter, connected by a wireless mesh, with continuous monitoring and automatic SMS or email alerts when termite-signature chewing was detected. GPS location included. Osseiran told phys.org in January 2012 that they expected to commercialize within twelve months.
Fourteen years later, the WiSPr is not on the market. No commercial IoT acoustic termite monitoring system is.
94.5% Accuracy, Zero Commercial Products
The July 2025 arXiv paper represents the state of the art. Its CNN-LSTM hybrid outperformed standalone CNN models and standalone LSTM models on the same dataset. Researchers collected acoustic data from termite-infested and clean wood samples under controlled conditions, extracted MFCC features, and trained the model to classify each sample. The 95.8% recall rate means the false-negative rate is low, which is the metric that matters most for a destructive pest: the cost of missing an active infestation vastly exceeds the cost of a false alarm.
An earlier study, published in IOP Conference Series: Earth and Environmental Science, achieved 93.2% accuracy using a simpler Support Vector Machine classifier on time-domain features, specifically energy and entropy extracted directly from raw acoustic signals without frequency transformation. That approach has the advantage of lower computational cost, potentially running on a microcontroller embedded in a monitoring station.
Both of these models work. Neither is available as a product you can buy, deploy, or spec into a construction project.
Why the Gap Persists
The economics of pest control work against early detection. Companies like Terminix (now Rentokil Terminix) and Orkin generate revenue from treatment contracts and annual protection plans, not from detection technology. Their incentive is to find termites on an initial inspection, sell a treatment, and lock the homeowner into a monitoring contract that guarantees repeat revenue. An IoT sensor array that detects infestations before they require professional treatment would cannibalize the highest-margin part of their business.
Insurance companies have no stake either. Termite damage is explicitly excluded from standard homeowner policies, which means insurers bear no cost from missed infestations. They have no financial reason to fund, develop, or require better detection technology.
Regulatory bodies set the inspection standards, and those standards describe methodology, not technology. NPMA's WDI Inspection Report (NPMA-33) requires the inspector to report visible evidence of wood-destroying insects and conditions conducive to infestation. It does not require acoustic sensing, thermal imaging, microwave radar, or any instrument beyond the inspector's eyes and a probing tool. State licensing boards for pest control operators test applicants on identification, treatment chemistry, and safety protocols. No state examination includes acoustic signal analysis or machine learning output interpretation.
Research-to-product pipelines for construction technology are punishingly slow even when incentives align. When they do not, the pipeline stalls entirely. WiSPr stalled. No startup has picked up the CNN-LSTM approach. Working technology sits in a paper. That paper sits on arXiv. And the termites keep eating.
What It Would Take
The hardware is not expensive. A MEMS microphone suitable for acoustic termite detection costs under $2 in volume. A low-power microcontroller capable of running an SVM classifier on time-domain features costs under $5. A battery-powered wireless sensor node with a five-year lifespan, comparable to existing smart home sensors from companies like Aqara or Eve, could plausibly retail for $30 to $50 per unit. Twenty units around a home's foundation would cost $600 to $1,000, roughly the same as a single spot treatment for a localized subterranean infestation.
Environmental robustness is the harder problem. Lab conditions are quiet. A real home has HVAC systems cycling, plumbing flowing, refrigerators compressing, kids running, garage doors opening. That CNN-LSTM model was trained on controlled data. Deploying it in a 1,800-square-foot ranch in Houston with two dogs and a teenager requires a noise-rejection pipeline that the current research does not address. This is an engineering problem, not a science problem. Speech recognition solved analogous noise-robustness challenges a decade ago. Transfer of those techniques to acoustic pest detection is straightforward but requires funding and product-development effort that nobody is spending.
Regulatory action could accelerate adoption. If HUD required acoustic or electronic verification for WDI reports on FHA-insured properties, the market would respond within eighteen months. If the VA added a technology requirement to its pest inspection mandate, vendors would build to the spec. But federal housing agencies move slowly, and pest inspection standards rank well below lead paint, radon, and structural safety on their priority list.
What to Do If You Are Buying a Home in Termite Country
Ask your inspector whether they use any electronic detection equipment. Most will say no. Some will mention moisture meters, which detect conditions conducive to subterranean termites but do not detect termites themselves. A few will own a Termatrac or acoustic emission device. If your inspector uses one, ask how they interpret the output and how many hours of training they have on it. If the answer is vague, it is not meaningfully better than the screwdriver.
Request that the inspector check areas beyond the standard walkthrough: inside finished basement walls, behind built-in cabinetry, above drop ceilings, and around bath traps. These are high-probability locations for hidden infestations that a twenty-minute visual sweep routinely misses.
Budget for annual inspections after purchase, not just the pre-closing check. NPMA recommends yearly inspections in high-risk zones. Cost: $100 to $300 per visit, a fraction of the average $3,000 to $10,000 repair bill for a caught-late infestation, and a rounding error against the $70,000 bill for severe structural damage that was invisible until the floor gave way.
And watch the IoT market. Continuous acoustic monitoring with AI classification is technically ready. What is missing is the business model and the regulatory push. When those arrive, and the cost delta between a $200 annual inspection and a $600 permanent sensor array becomes obvious to the first insurance company willing to offer a termite endorsement, the industry will move fast. Until then, the neural network sits in a paper, the termites sit in your walls, and the inspector sits on a ten-percent detection rate that everybody in the industry knows about and nobody talks about.