An auditory-based scene change detection in audio data

The problem of auditory scene segmentation plays important role in many audio analysis and processing tasks. The accuracy and robustness of segmentation step have influence on the remaining stages in audio processing chain. In the work a dedicated system for segmentation of the audio stream is presented. The segmentation scheme uses Delta-BIC and metric-based techniques to determine the change-point in audio data. For this purpose a dedicated auditory feature has been proposed, which is based on the gammatone filter bank. The proposed feature (GTEAD) has been designed using inter-channel analysis of the auditory filter bank outputs. For each channel, the temporal envelope and its periodic self-similarities have been calculated. Then, the distances between obtained signals from the neighbouring channels have been computed resulting in the final feature vector. The performance of the GTEAD feature has been compared to the popular MFCC feature using database of audio streams with defined single change-point in each example. The obtained results show that GTEAD feature outperforms MFCC feature in terms of accuracy and the number of detected points.