Audio lab · Sound event detection
Sound event detection: interactive demo
Sound event detection finds when target sounds occur in a stream of audio and marks where each one starts and ends. This demo runs a simple energy detector over an eight-second synthetic scene of knocks, beeps and low-frequency rumbles and scores it against ground truth, so you can see how the feature and the threshold trade hits against false alarms.
These demos are illustrative. Every signal is synthetic, generated and processed in your browser, and every number shown is computed from those signals. They show the principles, not the measured performance of a product.
Press a listen button to play a short sample. Sound starts only when you press a button. No microphone or uploads needed.
Sound event detection / edge audio
Catch the events. Skip the rumble.
What this shows: the same threshold detector gives very different hits and false alarms depending on the feature it watches.
Tune the detector
Lower thresholds catch quiet events and risk false alarms.
Eight targets (four knocks, four beeps) at different levels, plus two low-frequency rumbles that should not trigger.
What the detector sees
32 ms frames · 16 ms hopGreen: target foundHollow: target missedGrey: rumble (not a target)Blue: detector feature and correct detectionRed: threshold and false alarm
Always-on detectors on phones, wearables and sensors, often as the cheap first stage before a model. See on-device audio AI.
What this experiment models
An 8-second, 16 kHz synthetic scene: pink background noise at the chosen RMS level (dBFS before playback volume), four damped multi-resonance knocks, four 2.6 kHz beeps and two band-limited low-frequency rumbles. Each 512-sample Hann-windowed frame (16 ms hop) gives a log energy, either over the whole band or over 1–6 kHz. The floor is the median over the clip. Frames above floor + threshold are merged (gaps up to 3 frames, minimum 2 frames). A target counts as detected if any detection overlaps it; a detection that overlaps no target is a false alarm. This is a classical energy detector, not a trained model; learned detectors replace the feature and threshold, but the hit/false-alarm tradeoff remains.
The principle
How it works
The detector splits the audio into short frames, computes a log energy for each, and flags frames that rise a set number of decibels above the clip’s median level. Neighbouring flagged frames are merged into events. Watching broadband energy, the rumbles look like events and trigger false alarms; restricting the feature to the 1–6 kHz band, where the knocks and beeps carry their energy, ignores them. A lower threshold catches quieter targets and lets more noise through; a higher one does the reverse.
This trade-off is the central design decision for always-on detectors on phones, wearables and sensors. A cheap energy or band-energy stage often runs first and wakes a trained classifier only when something happens, which keeps power low. Learned detectors replace the hand-picked feature and threshold, but the trade-off remains: it moves into the choice of operating point. See on-device audio ai.
Audio lab
More demos
All six demos on one page: the audio labOn-device audio AI →