Abstracts Submitted: 469
Number of Users: 532
View Abstracts Submitted
Back to home Page
The situation monitoring system with camera has been widely used for the safety of people recently. However, the development of the systems is limited by the scope of monitors and the problem of privacy. It is considered that indoor environmental sound classification would be more practical in these systems than camera. In this paper, we address the disadvantage of the performance of indoor environmental sound classification would be decreased when a period of sound which was not used at construction of the acoustic model, the unknown class sound, is inputted to the acoustic model. As the conventional method, the unknown class sound inputted to the acoustic model would be classified as one of the known classes forcibly by learned acoustic model. To solve this problem, we proposed a method to classify indoor environmental sound based on self-generated acoustic model. In the proposed method, input sound would be discriminated between known and unknown, then the acoustic model will be reconstructed while the sound is classified into unknown class. In other words, the unknown class sound would be seemed as a new class added into the acoustic model. In this research, as the acoustic feature to be inputted to the DNN acoustic model, high dimensional acoustic feature having the highest classification performance in the previous research is used. We proposed a method which considered the output of deep neural network (DNN) acoustic model as output probability distribution. The difference between template distribution which is obtained by inputting known class sound and output probability distribution would be utilized for known-unknown sound classification. The scale of the difference between two probability distributions is based on Kullback Leibler divergence (KL divergence). Meanwhile, a sample set expansion method which is proposed by X. Song, X. Gao, Y. Ding and Z. Wang would be used to reconstruct the acoustic model based on an unknown class sound. An evaluation experiment has been conducted to prove the effectiveness of the self-generated acoustic model which we proposed. The results of the experiment have shown that the proposed method can obtain higher performance of indoor environmental sound classification than the conventional method.
© Copyright 2017 All Rights Reserved