Acoustic Drone Detection
- Osinto HQ
- 2 hours ago
- 7 min read
Drones represent a complex and evolving threat to both civilian and military infrastructure globally. In recent years Unmanned Aircraft System (UAS) operators have demonstrated their ability to seemingly surveil and attack targets around the world at will.
In the last two years small drones have notably been used to help overthrow the Assad regime in Syria [Reuters], take out military helicopters in Colombia [BBC] and bolster the attack capabilities of military junta in Myanmar [Reuters] to name but a few examples.
Nowhere however do drones dominate the skies more than above the battlefields of Ukraine of course, a fact that will have escaped few people's attention.
That these small aircraft are often assembled from Commercial-Off-The-Shelf (COTS) parts makes their further proliferation a near certainty. Furthermore the wide availability of components aids plausible deniability for operators, this very ambiguity making them an attractive tool for grey zone / hybrid warfare operations - making small UAVs all the more attractive to a range of state and non-state actors alike.
C-UAS Passive Acoustic Monitoring (PAM) aka Shazam for drones
Passive Acoustic Monitoring (PAM) systems are an increasingly popular method for detecting drones. For an overview of the technological components of acoustic monitoring systems see our post Acoustic Detection - from Conservation to Conflict Zones.
In Counter-UAS (C-UAS) applications acoustic detectors typically consist of an array of microphones (or particle velocity probes) used to detect the presence of drones. Most incorporate neural network classification algorithms to match detected audio signatures to libraries of known threats along with % confidence scores that a given smaple belongs to a particular class.
High levels of activity and interest in C-UAS applications of acoustic detection are being driven by a variety of factors:
Low-cost - commercially available parts keep costs low enough for wide area coverage (SkyFortress nodes - top picture - cost $400-1,000 USD / unit with 14,000 deployed and NATO funding a further 15,000 [United24] in Ukraine)
Low / no Radio Frequency (RF) emissions - make acoustic networks resilient to enemy targeting in contrast to systems like radar for example, whose active emissions paint them as targets on the battlefield
Battle-proven - in Ukraine as a viable defence against modern and evolving drone threats (Shahed-136 / Geran class one way attack / striek drones in particular) including those operated by fibre optic control and able to navigate autonomously (devoid of their own RF emissions signature)
Border / perimeter security applicable - used as part of layered air defence for frontier monitoring / early warning eg. to cue more expensive, higher fidelity electro-optical/ radar systems or civilian air raid alert systems
Proven in Ukraine, in demand across NATO
“They get very accurate information…sent out to all the fire teams who have an iPad showing the route of flight of these one-way UAVs coming in.”
Gen. James Hecker, Commander U.S. Air Forces Europe, Commander NATO Allied Air Command [Aviation Week].
Interest for Ukraine’s acoustic detection systems was growing even before recent NATO airspace incursions:
The US expressed interest in Ukrainian acoustic detection systems for low cost air defence applications as far back as 2024 [The War Zone]
The Lithuanian Armed Forces reportedly began testing a Ukrainian acoustic air threat detection system (rumoured to be Sky Fortress) in 2025 for deployment alongside existing radar systems in 2026 [Militarnyi]
Ukraine’s Sky Fortress was demonstrated to NATO militaries at Ramstein Air Base in Germany in 2024 [Defense One]
Pictured below is one of the more recent developments in Ukraine's arsenal of acoustic detection devices - the Zvook tactical portable sensor:
“The U.S. should integrate a low-cost acoustic network to detect aerial threats developed by Ukraine into its own air defense systems”
Commanding General of U.S. Army Space and Missile Defense Command, July 2024 [TWZ].
Acoustic drone detection systems overview
An increasing number of companies across Europe are known to be developing counter drone acoustic systems. Examples below show a braod range of acoustic node designs being tested, including:
Reported unit costs in the £6.5k to £12k range
Participants in NATO's DIANA accelerator
Founding team includes former French special forces operator Louis Saillans
BeephoniX 🇳🇱
Radboud University biophysics department spin-out
Komodo systems deployed in Ukraine with reported 0-5km range | Tetrahedral Iguana (pictured) demonstrated to Danish military September '25, reportedly also used for artillery, mortar and rifle fire ranging
Fenek 🇺🇦
Offer both air and border defence variants
Established collaboration with Dutch military, vehicle integrations, used for shot ranging applications as well as drone detection | Only company other than Welles Acoustics found utilising particle velocity sensor technology
Monava 🇸🇪
Claim TRL6, paid pilot programme with unspecified defence contractor and first units sold
Squarehead 🇳🇴
Appears an evolution of Norsonic's hextile acoustic camera (retail @ £25k+ / unit)
Welles Acoustics 🇵🇱 (acquired by Quantum Systems 02/25)
Polish particle velocity sensors, already integrated onto Quantum Systems' ISR drone platform, also used eg. for airborne artillery ranging
Zvook / SkyFortress 🇺🇦
Market leading Ukrainian systems, deployed at scale
Machine Learning (ML) drone classification
Most Passive Acoustic Monitoring (PAM) systems make use of neural networks for classifying detected sounds:
Convolution Neural Network (CNN) architectures appear to provide best results for classification
ML classifiers perform analysis on a spectrogram of recorded audio, typically transformed into Mel-Frequency Cepstral Coefficients (MFCCs) - the models themselves are often computer vision models that perform the categoristaion on the visual spectrogram rather than the audio input itself
Availability and quality of data used to train these neural nets is of paramount importance, directly correlating to accuracy of detections
“Recent advances in deep artificial neural networks (DNNs) have transformed the field of machine learning, frequently outperforming traditional signal processing techniques in the domain of acoustic event detection and classification.”
Kahl et al | Cornell University, Technische Universität Chemnitz, 2020 [BirdNET: A deep learning solution for avian diversity monitoring]
Mia Y. Wang - Drone Audio-Visualization Tool, The Drone Lab at The College of Charleston - showing a mel spectrogram of a DJI Matric 600P drone
There appears to be a lack of labelled training data for acoustic drone detection systems
“…the lack of acoustic drone datasets restricts the ability to implement an effective solution using deep learning algorithms”
[Al Emadi | DDI:Drones Detection and Identification using Deep Learning | Qatar University, Jan ’21]
“While our study presents a substantial foundation, acknowledging the need for further validation across diverse UAV platforms, environmental conditions, and real-world deployment challenges is imperative for realizing the full potential of UAV audio analysis”
[Wang et al | College of Charleston, Anhui Polytechnic University, Purdue University 2024]
Public datasets uncovered during this research include:
One of the larger public datasets available today is 'drone-audio-detection-samples' available here on HuggingFace.
Wang et al: A Multiclass Acoustic Dataset and Interactive Tool for Analyzing Drone Signatures in Real-World Environments | College of Charleston, September 2025
Svanström et al: A dataset for multi-sensor drone detection (pictured below) recorded at single location, max. 200m distance from drones, includes noise from camera mount as pulled from video feed | Produced in partnership with Swedish Armed Forces | Used by Helsing in Defence Tech hackathons in 2025 for acoustic drone detection challenges set for participants
Al-Emadi et al: Audio-Based Drone Detection and Identification Using Deep Learning Techniques with Dataset Enhancement through Generative Adversarial Networks | Dataset on Github - all recorded indoors
It's hard to ascertain the extent to which a lack of sufficiently well labelled acoustic data is a genuine impediment to the development of drone detection systems. It may simply be the case that system developers are maintaining good Operational Security (OPSEC) and not divulging details of proprietary data in a sensitive domain.
Regardless, the Ukrainian armed forces are highly likely to have a data advantage today, albeit one skewed towards a Russian threat profile. It would therefore seem pertinent for European players to fund / partner / collaborate with Ukrainian entities on acoustic drone detection in the near-term. The pragmatic new Ukrainian Minister of Defence Mykhailo Federov has in fact just unveiled the Brave1 Dataroom initiative which aims to provide access to just this sort of data to allied partners.
The training data a nerual network classifier is trained upon should match the acoustic environment in which a network is to be deployed as closely as possible. A range of synthetic data augmentation techniques (eg. adding environmental sounds, different signal:noise ratios etc) help facilitate this.
Google DeepMind's bioacoustic model - transferable learnings?
Google DeepMind's August '25 research paper 'Perch 2.0: The Bittern Lesson for Bioacoustics' yields some important insights for bioacoustic monitoring in ecological and conservation contexts. We hypothesise cross-domain applicability in the acoustic detection of drones:
Relatively small models (eg. a 12 million parameter EfficientNet-B3 derived Convolutional Neural Network (CNN) vs leading Large Language Models which commonly feature several billions of parameters) perform well at acoustic classification tasks and can run on computationally modest (and hence power efficient) hardware
The best performing Machine Learning (ML) models on general audio problems today are those that use supervised learning - training using labelled data
“Supervised pre-training benefits particularly from having fine-grained labels” - more granular labelling of training data results in improved model performance
Dataset diversity (eg. focused vs soundscape) and subsequent augmentation are important
Recommendations for acoustic drone detection system development and deployment
Deploy low-cost sensors based on commercial-off-the-shelf technologies, fast
Node hardware should be robust, simple, widely available and ideally have a sovereign / allied supply chain
As a relatively blunt instrument we'd be tempted to use widely spaced nodes for coarse triangulation of signals rather than focusing on the development of complex individual nodes with multiple sensors aiming for high fidelity geolocation - other sensor modalities (eg. electro-optical, RF, radar) do this better
Focus funds and resources on data collection, labelling and ML model development - building a data flywheel that feeds back into iterative ML model development
Significant effort should go into collating extensive training data featuring fine grained labels, significant augmentation and a mix of focused, soundscape and negative audio examples for the model to learn
Prototyping, research and development
Our own recent research and development work in the field of acoustic classification has shown that:
Nodes with acceptable performance can be produced in Europe at sub $1k / unit pricing
It should be feasible to curate a training dataset for £100k - £750k within a 60-90 day timeframe
CPU inference for acoustic classifiers is feasible given modest model parameter counts (eg. we had a Convolution Neural Network (CNN) classifier prototype running on an ultra lightweight Raspberry Pi Zero2W)
Low compute requirements mean low power requirements, which is helpful for edge deployments in austere / remote environments where solar PV and li-ion batteries can suffice to power a system
If you'd like to discuss any of the issues raised in this research, or should you have market intelligence, research and development or prototyping requirements of your own, please don't hesitate to contact us here or via hello@osinto.com.