Project Murmur
Inspiration
Recent conflicts in the Middle East have shown how drones have become a persistent threat to civilians, infrastructure, shipping, and military facilities. Drone attacks can arrive with limited warning and may operate alongside missiles or other aerial threats; recent attacks in Dubai, for example, involved multiple drones and affected both military and civilian areas.
Early detection matters because additional warning time can enable people to seek shelter, avoid exposed areas, and make better-informed safety decisions. Murmur explores whether distributed mobile devices can contribute to that early-warning picture by detecting acoustic and wireless signatures locally and sharing observations even when conventional connectivity is unavailable. Drone detection normally relies on specialized infrastructure such as radar, dedicated RF receivers, or acoustic arrays. But in dense urban environments, something much more interesting already exists: thousands of mobile phones equipped with microphones, Bluetooth radios, cameras, location sensors, and powerful on-device compute.
We asked:
What if those phones could work together as a distributed early-warning network?
Murmur explores whether mobile phones can detect patterns associated with nearby drone activity through acoustic and accessible wireless signatures, then share compact observations across an offline mesh network.
The goal is not just detection. When seconds matter, Murmur aims to transform distributed observations into understandable situational intelligence that can help people identify potential threats and move toward designated safer areas.
What it does
Murmur turns mobile phones into distributed situational-awareness nodes.
Each phone passively observes its environment using acoustic sensing and accessible wireless telemetry such as Bluetooth activity.
Instead of treating a single observation as proof of a drone, Murmur combines multiple signals over time and produces structured SensorEvent messages.
A typical event might contain:
event_type
timestamp
approximate_location
acoustic_anomaly_score
wireless_anomaly_score
confidence
duration
These compact events are designed to travel across a Meshtastic-based LoRa mesh, allowing observations to propagate between nodes even when Internet or cellular connectivity is unavailable.
Phone A
│
▼
Meshtastic Node A
│
│ LoRa
▼
Meshtastic Node B
│
▼
Phone B
Multiple observations can then be correlated across time and location.
Rather than presenting raw sensor data, Murmur's AI harness turns these events into understandable information such as:
- What was observed?
- Where was it observed?
- How long has it persisted?
- Which sensors contributed evidence?
- How strong is the evidence?
- Are nearby nodes observing related activity?
- What relevant safety information is available?
How we built it
Murmur follows an edge-first architecture:
Sensors
↓
Feature Extraction
↓
Specialized ML
↓
Sensor Fusion
↓
SensorEvent
↓
Offline Mesh
↓
AI Harness
↓
Human
Edge sensing
On Apple devices, we use native frameworks to collect permitted sensor information including:
- Audio
- Bluetooth
- Location
- Motion
- Potentially computer vision
Acoustic signals are transformed into frequency and temporal features using techniques such as:
- FFT
- STFT
- Spectrogram analysis
- Spectral energy
- Periodicity
- Temporal features
Bluetooth observations provide features such as:
- Advertisement activity
- RSSI
- Persistence
- Appearance/disappearance patterns
- Advertisement timing
These are treated as experimental wireless observations rather than arbitrary raw RF measurements.
An iPhone is not a software-defined radio, so Murmur does not assume access to raw RF spectrum.
Specialized machine learning
Specialized ML models handle signal classification and anomaly detection.
The principle is:
Physical signal
↓
Feature extraction
↓
Specialized ML
↓
Structured evidence
A Bluetooth anomaly, acoustic anomaly, or visual observation is not independently treated as proof of a drone.
Instead, Murmur investigates whether these observations contain complementary information that becomes more useful when correlated.
Apple Foundation Models AI Harness
Above the detection system sits our AI harness powered by Apple Foundation Models.
The Foundation Model is intentionally not responsible for raw signal detection.
Instead, the AI harness provides controlled access to structured evidence and application tools.
For example:
Foundation Model
│
├── getRecentEvents()
├── getNearbyEvents()
├── compareNodeObservations()
├── getAcousticEvidence()
├── getWirelessEvidence()
└── getSafetyInformation()
This allows the model to:
- Explain recent events
- Compare observations between nodes
- Summarize incidents
- Explain uncertainty
- Query historical observations
- Generate structured reports
- Surface relevant safety information
The architectural boundary is simple:
Specialized ML measures the environment. Apple Foundation Models helps humans understand it.
Offline communication
Murmur proposes a Meshtastic-based LoRa network for resilient offline communication.
The phone communicates with a nearby Meshtastic-compatible radio node:
iPhone
│
Bluetooth
│
▼
LoRa Node
│
│ LoRa Mesh
▼
LoRa Node
│
Bluetooth
│
▼
iPhone
LoRa bandwidth is limited, so Murmur does not attempt to transmit:
❌ Continuous audio
❌ Video
❌ Camera frames
❌ Large sensor logs
Instead, devices exchange compact structured events:
✓ SensorEvent
✓ Timestamp
✓ Approximate location
✓ Event type
✓ Confidence
✓ Evidence summary
The principle is:
Move intelligence, not raw surveillance data.
Challenges we ran into
The biggest challenge was understanding what mobile hardware can actually observe.
An iPhone is not a software-defined radio, and normal applications cannot arbitrarily inspect raw Wi-Fi or cellular RF spectrum.
We therefore reframed the problem from:
"Can we detect RF using an iPhone?"
to:
Can legitimately accessible mobile sensors collectively provide enough information to distinguish drone-related activity from normal environmental activity?
False positives are another major challenge.
Bluetooth devices are everywhere, while:
- Aircraft
- Helicopters
- Vehicles
- Machinery
- Fans
- Power tools
- Crowds
- Urban noise
can produce observations that overlap with drone activity.
This means Murmur cannot depend on a single sensor.
The system needs:
Multimodal evidence
+
Temporal correlation
+
Spatial correlation
+
Multiple nodes
↓
Situational intelligence
Offline networking creates another constraint: LoRa bandwidth is extremely limited.
This led us to the SensorEvent architecture.
Raw sensor information stays on the device whenever possible. Only compact derived intelligence needs to travel through the network.
Finally, integrating generative AI required us to establish an evidence boundary.
Apple Foundation Models can explain and reason over information supplied by our AI harness, but it must never fabricate sensor evidence that the underlying detection pipeline did not produce.
Accomplishments that we're proud of
We're proud that Murmur became more than an AI drone detector.
We designed the architecture around three independent layers:
1. Edge Perception
Phones observe their local environment.
Audio
Bluetooth
Vision
Location
Motion
↓
Local ML
2. Offline Collective Intelligence
Compact observations can propagate through a Meshtastic network.
Node A ───── Node B
│ │
│ ├──── Node C
│ │
└────────────┴──── Node D
3. On-device Human Intelligence
Our Apple Foundation Models-powered AI harness converts structured observations into understandable situational information.
SensorEvents
↓
AI Harness
↓
Apple Foundation Models
↓
Human-readable intelligence
This means the architecture does not require continuous cloud connectivity to perform its core functions.
We're particularly proud of the privacy model.
Rather than creating a network that continuously uploads microphones and cameras to a central server, Murmur is designed to process sensitive sensor information locally and communicate derived events.
Move intelligence, not raw surveillance data.
What we learned
Our biggest lesson was:
Observation is not identification.
A Bluetooth anomaly is not a drone.
An acoustic anomaly is not a drone.
A visual anomaly is not necessarily a drone.
Each is simply a piece of evidence.
The interesting ML problem is determining whether multiple independent observations contain complementary information.
Wireless ────────┐
│
Acoustic ────────┤
│
Vision ──────────┼──→ Sensor Fusion
│
Context ─────────┤
│
Other Nodes ─────┘
↓
Candidate Event
We also learned that distributed intelligence changes the problem.
One phone observing an unusual acoustic pattern may provide weak evidence.
Several geographically separated nodes observing correlated activity within a short time window potentially provide much richer information.
And we learned that specialized ML and Foundation Models solve fundamentally different problems.
Specialized models detect patterns.
Apple Foundation Models helps humans understand those patterns.
What's next for Murmur
Our immediate priority is building the dataset.
We plan to conduct controlled experiments across:
DRONE_ABSENT
DRONE_PRESENT
FALSE_POSITIVE
environments and measure the contribution of each sensing modality independently.
We will compare:
Acoustic only
↓
Wireless only
↓
Vision only
↓
Acoustic + Wireless
↓
Acoustic + Vision
↓
Full Multimodal
↓
Multimodal + Distributed Nodes
This allows us to determine scientifically whether each additional source of information actually improves detection.
The next stage is deploying validated models through Core ML, allowing inference to remain on-device.
We then plan to connect multiple Murmur nodes through Meshtastic and investigate spatial-temporal correlation: whether observations propagating between geographically separated nodes can provide useful information about the presence and movement of an environmental event.
Finally, we'll expand our Apple Foundation Models AI harness with tools for:
- Querying nearby events
- Comparing observations between nodes
- Generating incident summaries
- Explaining confidence and uncertainty
- Querying historical events
- Surfacing relevant safety information
Our larger vision is simple:
Phones sense locally. The mesh communicates offline. Specialized ML finds the signal. Apple Foundation Models makes it understandable.
Together, Murmur explores whether devices people already carry can become a decentralized, privacy-preserving situational-awareness network.
Industrial Use Case: Offline Acoustic Predictive Maintenance
Murmur can extend beyond drone detection into industrial acoustic monitoring and predictive maintenance.
Industrial machinery often reveals early signs of failure through subtle changes in sound and vibration. Motors, pumps, compressors, fans, bearings, generators, and HVAC systems can develop distinctive acoustic patterns before a fault becomes visually obvious or causes complete failure.
Murmur turns an iPhone into a portable edge diagnostic node.
Machine / HVAC System
↓
Microphone + Sensors
↓
Acoustic Feature Extraction
↓
On-device ML
↓
Fault / Anomaly Detection
↓
Murmur AI Harness
↓
Apple Foundation Models
↓
Technician
The device can capture short acoustic samples and extract features such as spectral energy, frequency peaks, harmonics, periodicity, spectral changes, and temporal patterns.
Specialized on-device ML models can then compare those characteristics against normal operating behaviour and identify potential anomalies.
Instead of sending recordings to the cloud, analysis happens locally.
Our Apple Foundation Models-powered AI harness can then translate the structured diagnostic evidence into useful information for a technician:
“Abnormal periodic acoustic activity detected near HVAC Fan Unit 4. The strongest deviation is around the rotating-frequency harmonics. Inspection of the fan and bearing assembly is recommended.”
The Foundation Model does not diagnose the machine directly from raw audio. Specialized signal-processing and ML components generate the evidence first; the AI harness makes that evidence understandable and actionable.
Why offline matters
Industrial environments frequently have areas where cloud-dependent AI performs poorly or cannot operate at all:
- Underground facilities
- Tunnels
- Mines
- Plant rooms
- Basements
- Remote infrastructure
- Confined spaces
- Large industrial sites with unreliable connectivity
- Facilities where operational data cannot leave the site
Because Murmur's core intelligence runs on-device, technicians can continue collecting observations, detecting anomalies, and interacting with the AI system without relying on cloud connectivity.
A Meshtastic-based network could extend this further.
Pump Station
│
Murmur Node
│
│ LoRa
▼
HVAC Room ───── Compressor Room
│ │
Murmur Node Murmur Node
│ │
└──── Offline ─────┘
Mesh
Instead of transmitting continuous recordings, nodes can exchange compact events such as:
{
"asset": "HVAC_FAN_04",
"event": "acoustic_anomaly",
"confidence": 0.91,
"severity": "inspect",
"timestamp": "2026-08-30T04:32:10Z"
}
This allows Murmur to provide distributed equipment awareness without requiring continuous Internet connectivity or transmitting raw industrial sensor data.
The same architecture that listens for unusual activity in the sky can therefore listen for unusual behaviour inside critical infrastructure.
Murmur brings AI to environments where the cloud cannot reliably reach.
Log in or sign up for Devpost to join the conversation.