Blind Detection and Classification Algorithm for Smart Audio Monitoring System

This paper presents an efficient method to classify an audio event automatically. The proposed method blindly detects audio events before feature extraction which can be considered a novelty. Once an audio event is detected; after that, every frame belongs to the event is passed through the feature extraction and classification stages. Mel-scaled filter bank energy features were extracted frame-wise. The extracted features were classified into eleven subsequent classes of audio events using deep neural network (DNN) technique. The proposed methodology was validated on the TUT DCASE-2016 task 2 database which is a synthesised office environment with a controlled level of background noises, i.e., 6dB, 0dB and -6dB SNR. The proposed methodology reported an accuracy of over 96% in detecting audio events blindly whereas classification accuracy is reported to be over 80%. Obtained results were compared with the state-of-art methodologies reported in the literature.