Audio Event Detection Based on Cross Correlation in Selected Frequency Bands of Spectrogram

Audio event detection (AED) systems have various applications in modern world. Examples of applications include security systems, urban management and automatic monitoring in smart cities, and online multimedia processing. The noise and background sound vary in an urban environment, so frequency domain and normalized features usually show better efficiency in AED systems. This work proposes a Mel spectrogram-based approach that uses the spectral characteristics of audio signals and cross-correlations to build a dictionary of effective spectrogram frequency bands and their patterns in different audio events. Initially, the proposed approach extracts the Mel spectrogram of audio input. In the next step, a mathematical-statistical analysis is used to specify the effective frequency bands of the spectrogram in each audio event. The pattern of selected frequency bands varies due to the type of event, which can effectively help decrease spectrogram size as input feature, reduce errors and increase the accuracy of different AED methods. The proposed approach was implemented on the URBAN-SED database, and its efficiency was compared against deep learning base state-of-the-art researches in the field. According to the results, about 50% of the frequency bands in the spectrum are useless and can be discarded in the training process of an AED system without any loss in terms of accuracy.