Abnormal sound recognition with audio feature combination and modified GMM

This paper presents a new abnormal situation recognition approach based on environmental sound analysis. Considering the characteristic of household abnormal sound, several audio features and their combination are evaluated through substantial experiments. Longer audio frame is utilized to improve feature extraction for environmental sound. During classification, Gaussian Mixture distribution is trained to model the audio pattern and an adjustment coefficient is introduced to recognition process. The experimental results show that the mixed feature combined with Mel Frequency Cepstrum Coefficient (MFCC) and short-term energy performs efficient and reliable recognition effect with satisfied computing time cost. The practical system implementing the proposed approach can be used in many occasions such as home security monitoring, abnormal situation alarm, and environmental recognition etc.