Capture
Upload a recording, or open the live monitor and stream a microphone in rolling two-second windows.
- WAV · MP3 · FLAC · OGG · M4A
- Up to 25 MB and 10 minutes
Acoustic event monitoring · 17 sound classes
Every sound,
on the record.
SonicSentinel listens to recordings and live microphones, recognises ten kinds of sound — from a gunshot or breaking glass to a person calling for help — and checks every result with two independent models before it raises an alert.
01Why sound
Cameras see what is in frame. Microphones hear around corners, through walls and in the dark. SonicSentinel turns that sound into events a person can act on.
02Pipeline
From microphone
to decision.
Clean
Mixed to mono, DC offset removed, resampled to 22.05 kHz, spectral-gate noise reduction, silence trimmed and the peak normalised.
- Same steps for training and prediction
Segment
Cut into two-second segments that overlap by one second, so a short event is never split in half. Quiet segments are skipped.
- 2 s windows · 1 s hop
Listen twice
A Python ensemble reads 193 acoustic features on the server. A Teachable Machine network reads a 43 × 232 spectrogram in your browser.
- Neither model sees the other’s answer
Compare
The two probability sets are fused and compared. The result is a match, a weak match, a disagreement or an uncertain result.
- = ≈ ≠ ?
Act
Per-class alert rules decide severity, confirmation and escalation. Anything uncertain goes to a person in the review queue.
- Every step is kept in the audit trail
03What it hears
Ten sounds.
Each has a shape.
Each card draws the typical spectrogram of its class — time left to right, pitch bottom to top. Hover a card to play its scan.
Machinery Fault
Abnormal machine noise: grinding, knocking, bearing squeal, belt slip or rattling.
Raises an alertGlass Breaking
Window or glass shattering, including the scattered shards that follow the impact.
Raises an alertAlarm or Siren
Fire alarms, intrusion alarms, emergency-vehicle sirens and beeping alert tones.
Raises an alertVehicle Horn
Car, truck, motorbike or bus horns.
Logged, no alertAnimal Sound
Dogs barking, cats, livestock, birds, insects, frogs and other animal vocalisations.
Logged, no alertGunshot
Firearm discharge: a sharp broadband impulse followed by an echo tail.
Raises an alertPanic Scream
High-pitched screams that indicate fear or distress.
Raises an alertAggression or Violent Conflict
Aggressive shouting, heated arguments and sounds of physical conflict.
Raises an alertPerson Asking for Help
Defined safety phrases only: "Help me", "Somebody help", "Please help", "Call for help", "Emergency".
Raises an alertDrone Sound
Multirotor drone (UAV) flying, hovering or passing: buzzing rotor tones with beating between the propellers.
Raises an alertDrilling or Grinder Sound
Power tools: electric or hammer drills, jackhammers, angle and bench grinders, impact drivers.
Logged, no alertFireworks
Firework shells, rockets and firecracker strings: whistles, bursts and crackling tails.
Raises an alertVehicle Backfire
Exhaust backfire or afterfire pops from a vehicle, usually with engine sound around them.
Logged, no alertExplosion
Blasts and explosions: a heavy low-frequency boom with a long rumbling tail, sometimes debris.
Raises an alertDoor Impact
Doors being slammed, kicked, rammed or banged on, and knocking on doors.
Raises an alertGraffiti Spray
Aerosol spray cans: the mixing-ball rattle of a shaken can and bursts of spray hiss.
Raises an alertBackground Noise
Ambient sound with no event of interest: traffic hum, wind, rain, office or crowd murmur.
Logged, no alert04Two models, one verdict
Two models listen.
Then they compare notes.
On the server
Python ensemble
193 acoustic features per segment — MFCCs, mel bands, spectral contrast, pitch, onsets and the energy envelope — classified by XGBoost, a neural network and gradient boosting voting together.
91.2% test accuracy
Verdict
=
Acceptable match
Same class, similar confidence. The alert rule decides what happens next.
In your browser
Teachable Machine network
A convolutional network in the Teachable Machine audio format reads a 43 × 232 spectrogram of the loudest second of each segment, running on TensorFlow.js.
90.8% test accuracy
- Acceptable matchSame class, confidences close together.
- Weak matchSame class, but one model is much less sure.
- DisagreementDifferent classes. A reviewer listens.
- UncertainBoth unsure. Logged as unknown and reviewed.
Bars above are an example segment. Accuracy is measured per recording on 1803 unseen test recordings.
17sound classes
12,018recordings in the dataset
193acoustic features
43×232spectrogram input
2second analysis window
05From detection to decision
An alert has
to be earned.
Every class has its own rule. Here is the gunshot rule, step by step.
- 00:00.0Heard twiceA gunshot is detected in two consecutive segments.
- 00:01.2Checked against the ruleBoth models agree, confidence at least 70 %, a clear margin over the next class, audio quality at least acceptable.
- 00:01.3Critical alertSecurity is notified with the recommended action: check nearby cameras, keep people away.
- 00:30.0EscalationThree shots within 30 seconds escalate to the security lead and emergency services.
- laterAcknowledged, reviewed, closedAn operator acknowledges it, a reviewer confirms the category, and the audit trail keeps every step.
PythonBackground noise
Teachable MachineBackground noise
DecisionNo event
06Live monitor
Open a microphone.
Watch it think.
The live monitor analyses a two-second window every second: waveform, spectrum, both models and the rule decision, with an alert toast and sound when something critical happens.
Try the live monitor07Responsible by design
Built for controlled,
ethical testing.
- A prototype, said plainlyNot a certified emergency-response or law-enforcement system.
- Events, not eavesdroppingLive monitoring stores audio only for windows with an event, unless an administrator changes it.
- Retention limitsAudio is deleted after 30 days, live audio after 7, records after a year. All configurable.
- RolesOperators, reviewers, maintenance and administrators each see only what their job needs.
- Audit trailUploads, predictions, reviews and rule changes are all written down.
08Questions
Asked often.
Which sounds can it recognise?
Machinery Fault, Glass Breaking, Alarm or Siren, Vehicle Horn, Animal Sound, Gunshot, Panic Scream, Aggression, Person Asking for Help, Drone Sound, Drilling or Grinder Sound, Fireworks, Vehicle Backfire, Explosion, Door Impact, Graffiti Spray, Background Noise. Anything that fits none of them is reported as an unknown sound.
Which files can I analyse?
WAV, MP3, FLAC, OGG and M4A, from half a second up to ten minutes and 25 MB. Batches of up to 50 files are accepted.
Where does my audio go?
Uploads are stored on this server so they can be analysed, played back and reviewed. The Teachable Machine model runs in your own browser. No third-party cloud service classifies the audio.
What happens when the two models disagree?
The event is marked as a disagreement and goes to the review queue, where a reviewer listens to the segment and confirms or corrects the category.
Can the models be retrained?
Yes. Both are trained on the same dataset and splits: python -m python_models.train for the Python model and python -m scripts.train_gtm_model for the Teachable Machine network. The app picks up new model files without a restart.
Put a microphone on it.
Open the consoleDemo accounts are listed on the sign-in page — for example admin / Admin@2026.