Skip to main content

Sound Event Detection

Sound Event Detection (SED) moves beyond simple "wake word" recognition to a broader understanding of the acoustic environment. By utilizing the Syntiant® NDP, devices can be trained to recognize non-speech signatures—such as glass breaking, alarms, or a baby crying—with the same power efficiency as a voice trigger. This allows for a proactive security and monitoring posture that remains functional even in total darkness.

  • Advanced Audio Detection (Glass Break & Alarms): Traditional glass break sensors often rely on simple shock or frequency thresholds, leading to high false alarm rates from kitchen clatter. The NDP uses deep learning to identify the specific temporal and spectral sequence of a glass break: the initial "thud" of the impact followed by the high-frequency "shimmer" of the shattering glass.

  • Safety Monitoring (Baby Cry & Alarm Detection): In smart nursery or elderly care applications, the NDP can distinguish a baby’s cry or a smoke alarm from general television noise or domestic chatter. This local processing ensures that the alert is instantaneous and does not require an active internet connection to function.

  • Acoustic Signature Mapping: To ensure the model is robust against environmental interference, we measure the Frequency Response (H(f)) of the microphone and housing to compensate for any structural dampening.

Formula: System Frequency Response (H(f)H(f))

H(f)=Y(f)X(f)H(f) = \frac{Y(f)}{X(f)}

(where X(f) is the input sound signal and Y(f) is the signal as captured by the NDP after passing through the device enclosure.)

  • Animal Sound & Domestic Pattern Recognition: The NDP120 series can be trained on specialized classes such as dog barking or water leakage sounds. By calculating the Spectral Centroid (SC), the system can distinguish between a high-pitched alarm and a low-frequency hum.

Formula: Spectral Centroid (SCSC)

SC=∑n=0N−1f(n)⋅x(n)∑n=0N−1x(n)SC = \frac{\sum_{n=0}^{N-1} f(n)\cdot x(n)}{\sum_{n=0}^{N-1} x(n)}

(where f(n) represents the frequency of bin n and x(n) represents the magnitude of that bin.)

⚠️💡The "Negative Dataset" Strategy

"When training, the 'Negative' data is as important as the 'Positive.' Include recordings of vacuum cleaners, dishwashers, and loud music in your training set. This teaches the NDP to treat these high-energy sounds as 'Background' so that it only consumes the full power of the host system when a true security event is detected."