From the Lab to the Field

Where the lab and the field part company

Each way a deployment differs from a training set has a cost, and most have a known mitigation.

Why 99% means less than it sounds

Near-perfect accuracy is common in the acoustic detection literature. Reviewing work on detecting whether a drone carries a payload, researchers at Université de Sherbrooke noted studies reporting around 99% accuracy, but with data from only one or two drone models recorded at up to 15 metres; one study that measured distance found accuracy falling to 70% at 100 metres, and wind noise above 54 dB significantly hurting predictions.Marcoux & Grondin A December 2025 review of the whole field in AIP Advances makes the same point about the data underneath: most public drone-sound datasets cover common consumer multirotors, record within tens of metres, and follow no shared standard for equipment or labelling, which makes fair comparison between models difficult.AIP Advances

The pattern is not unique to drones. A 2024 study of street-sound classifiers built into city streetlights reported 85.4% accuracy on a standard benchmark, then found in real deployments that the model worked well for most vehicles but had significant limitations for pedestrians and bicycles. The authors named the quality and quantity of training data as the main limitation.Discover Applied Sciences

A detector that has only heard its own microphone, its own field and its own drone has learned those as much as it has learned drones.

Familiar, unfamiliar and hardest

The Fraunhofer team trained a compact neural-network detector on 71 hours of drone recordings and 380 hours of background from their own eight-microphone array, plus public datasets, then tested it three ways. Crucially, they held out whole recording sessions, so the test never shared a day and place with training. They tuned every model to a fixed false-alarm rate on background audio and then measured how much drone sound it caught.Fraunhofer IDMT

Trained only on drone data, the detector caught 86% of drone segments on familiar recordings and 70% on public recordings made with other drones, microphones and sites. Pretraining it first on AudioSet, a large general collection of everyday sounds, lifted those to 92% and 78%; adding the full set of field-style augmentations brought the unfamiliar average to 83%. The hardest set, audio collected from the web, went from 17% to 66%.Fraunhofer IDMT

The authors attribute the remaining gaps to differences in drone types, recording hardware and post-processing, and suspect compression artefacts in web-sourced audio that their augmentations do not yet capture.Fraunhofer IDMT

Where field performance leaks away

Range is the gap that no training trick removes. On real flights recorded in Berne, Germany, the Fraunhofer detector caught 98–99% of drone sound within 50 metres at a 1% false-alarm rate, about 81% at 50–100 metres, 52% at 100–150 metres and almost nothing beyond 200 metres. Listening for two seconds instead of one raised those middle figures to 86% and 68%, because a longer window gives a faint, wavering signal more chance to stand out. The cost is a slower alert. The authors call the detector highly reliable to 100 metres and partly effective to about 150.Fraunhofer IDMT Denser spacing is the route DroSonic takes: many small networked nodes rather than one large array, so no drone has to be heard from far away.DroSonic

These ranges come from one microphone array, one site and a few camera drones (DJI Phantom, DJI Mavic and Align models).Fraunhofer IDMT Louder drones carry further, and quieter sites help, but the shape of the curve is the point: a single acoustic node covers a neighbourhood, not a region. Our explainer What a Microphone Can Hear sets out the physics of why.

The world is full of things that hum

An operator's tolerance for false alarms is far lower than a benchmark's. The Fraunhofer authors spell out what a 1% false-alarm rate means at the scale of individual one-second decisions: roughly 36 false segments an hour, which would be too many if each raised an alert. They argue that detections need merging over time and filtering by further logic, or confirmation by another sensor, with acoustics serving as a sensitive early cue.Fraunhofer IDMT

What fools a drone detector is telling. Tested on a public collection of 2,000 everyday sound clips, the model most often mistook washing machines, chainsaws and insects for drones, at 14–17% each, followed by vacuum cleaners at 9%. Every one is a steady mechanical or buzzing texture, like a rotor.Fraunhofer IDMT

Radar has its own version of this problem, with birds; our report Is It a Bird? covers it. The cure is the same in both cases: a second, independent signature before anyone acts.

The anechoic chamber has no wind

Laboratory recordings are often made in quiet rooms, or close to the drone on still days. Outdoors, wind blowing across the microphone is one of the strongest noise sources, and it sits in the same low frequencies as rotor tones. Physical protection helps a great deal: one outdoor study of drone-detection microphones found that a windscreen extended detection range in strong wind by roughly 31–131%, depending on the wind's direction.PubMed Central

Training for the world, testing like a buyer

The Fraunhofer results rank the fixes. Pretraining on a broad library of sounds was the single largest gain, on familiar and unfamiliar data alike. Augmentation came second and mattered most on mismatched data: simulating different microphone responses gave the biggest single jump, then mixing in background noise, masking parts of the spectrum, and pitch-shifting drone sound to mimic other rotor speeds.Fraunhofer IDMT Augmentation is not free: the most aggressively augmented model raised false alarms on unfamiliar everyday sounds slightly, from about 2% to 2.5% at the 1% setting.Fraunhofer IDMT

For a buyer, the questions follow directly:

1. Was the test data recorded on different days, places and microphones from the training data? If not, the result is a lab result.
2. What was the false-alarm rate when the detection rate was measured? A detection rate without one is meaningless.
3. How does detection fall with distance? Ask for results in distance bands, not an average.
4. Which sounds cause false alarms, and what happens next? Look for event logic over time and a second sensor before action.
5. Which drones were tested? Size, payload and propulsion change the sound.

Field data over benchmark scores

Every row of the table above applies to DroSonic's own sensors, which is why its development moved from recordings to open ground early. Detection runs as neural networks on low-cost edge hardware inside each node, so the model, the microphone and the site are tested together rather than one at a time.DroSonic

In September 2026 a prototype was tested against live drones at DIRT Days 26.04, with wind, vehicles and an electronic environment built to resemble a contested battlefield; DroSonic reports that it heard drones other sensors missed.DroSonic That is an account, not a published benchmark. The test this report recommends, detection by distance at a fixed false-alarm rate on unseen sites, is the standard any buyer should hold acoustic detection to, DroSonic's included.