Research On The Classification Of Indoor Multi-channel Human Activities Sound Events

With the development of information technology,there are more and more ways of human-computer interaction.The realization of sound event recognition technology which is the key technology to assist human-computer interaction gradually become research hotspot.At present,the sound event recognition algorithm is aiming at public places,and there is no specific optimization for indoor environment sound characteristics,and the current algorithm often uses mono audio without using a microphone array.In addition,the current stage of sound event recognition performance is poor,and the robustness is not good.In view of these problems,this paper studies the classification of indoor multi-channel human activity sound events in complex environments.The main work of this paper is as follows:1.The traditional sound event recognition algorithm is studied,and the sound collection,preprocessing,feature extraction and classifier are studied respectively,and the performance is analyzed through simulation.2.Aiming at the reverberation sound conditions in indoor complex environment,a beamforming algorithm suitable for reverberation conditions is studied.This method is based on blind source multi-input and output impulse response shortening algorithm,which can effectively improve reverberation in indoor environment.The time and frequency caused are blurred.For the shortcomings of the indoor sound event signal,it is difficult to obtain the accurate angle of arrival and the geometric parameters of different microphone arrays.The minimum variance and distortion-free response beamforming algorithm and complex Gaussian mixture model are studied.The algorithm can be widely applied to linear microphone arrays with different parameters.In response to multi-source conditions,as well as improving the signal-to-noise ratio of sound event signals.3.Feature extraction of indoor sound event signals,in addition to extracting traditional Mel Cepstrum coefficients,gamma pitch filter Cepstrum coefficient characteristics,extracting the angular spectrum of the signal and frequency screening spectrogram characteristics for the indoor environment.4.This paper analyzes the voice recognition method based on convolutional neural network(CNN)and recurrent neural network(RNN),and proposes to identify indoor sound event signals based on convolutional recurrent neural network(CRNN).Compared with traditional classifiers,it has the advantages of high recognition performance and strong robustness.Because the indoor sound event data set is large,the convolutional recurrent neural network model is relatively complex.For the model over-combination phenomenon,this paper also studies the Batch Normalization and Dropout anti-overfiting optimization algorithms.The effects of different characteristics,different neural network parameters and different neural network institutions on the classification results were analyzed by actual experiments.After the parameters are tuned,the recognition performance of the convolutional neural network can reach 98% of the F1 score,which verifies the effectiveness of the proposed algorithm.