Sound Detection,classification And Localization Under Noise Conditions

Acoustic event detection refers to the process of detecting the segments with clear semantics in the continuous audio signal.It is the basis for the recognition and semantic understanding of the surrounding acoustic environment by the machine,and has important significance in the recognition and perception of the acoustic environment by artificial intelligence.This article embarks from the development process in the field of acoustic event detection,acoustic event detection field in detail at each stage the main problem,research direction and related achievements,review of the history of acoustic event detection field,on the basis of the classical acoustic event detection method to analyze its advantages and disadvantages,and further put forward in this paper,the method of model,It also introduces the problems that the method of this paper aims to deal with,the main ideas and the final results.Aiming at the problem that the dynamic routing algorithm in capsule network has super parameters and the iteration result is affected by the data balance,this paper proposes a dynamic routing algorithm based on Gaussian mixture model and applies it to capsule network.The improved dynamic routing algorithm does not need to artificially set the number of iterations as the super parameter,and the interpretation of the iterative process is better than the original dynamic routing algorithm.The model proposed in this paper is mainly applied in noisy environment,and the noise energy may be much higher than the acoustic event signal energy that needs to be detected.Meanwhile,the source of noise signal is not just white Gaussian noise,but also may come from other acoustic events.In the case of uncertain noise distribution,neither simple wavelet transform denoising nor simple Kalman filtering can meet the requirements of denoising,so it is necessary to combine the two methods.In view of this situation,this paper adopts a denoising method based on wavelet and Kalman filter.By combining the advantages of wavelet transform multi-scale multi-resolution and Kalman filter with the minimum mean square error,the performance of the denoising algorithm is effectively improved.Experimental results show that Caps Net-LSTM model and Caps-GMM-LSTM model both achieve more than 90% recognition accuracy under the SNR of 0d B,and are higher than Capsule network and other classical models under all SNR conditions.At the same time,the experiment also found that with the gradual reduction of SNR,the accuracy of Caps Net-LSTM model and Caps-GMM-LSTM model also decreased significantly,indicating the necessity of signal noise reduction.