Providing an Adaptive Model with two Adjustable Parameters for Audio Event Detection and Classification in Environmental Signals

Audio event detection (AED) is a modern way to collect data about human activities in the workplace or in other life environments. We proposed a novel adaptable model based on using two parameters, α and ᵦ to detect all audio events that may be present in a given record accompanied by their time limits in which they occur. After feature extraction and setting the values of the two key parameters, alpha and beta, the audio sequence will be sent into two distinct sub-systems for event detection. The outputs from the two sub-classifiers are then combined and necessary refinements are made on the event time limits. The final detected events are sent to the KNN classifier. The parameters serve as a trade-off tool between precision and recall expectation in the detection process. In the tests, 16 different audio events of an office room were detected, some being similar to each other and some have very similar characteristics to those of the background noise. At frame-based (FB) level, the precision rate was 70.1%, the rate of recall was 75.8%, and F1-measure was 72.8%. The F1-measure has increased by 10.8% suggesting promising applications of the model.