Research On Acoustic Event Detection Method Based On CRNN-HMM

Sounds carry a large amount of information about our everyday environment and physical events that take place in it.Based on audio content search,speech recognition,robotics,driverless cars,and intelligent surveillance systems,acoustic information is used to identify activities in their environment.At present,non-speech understanding and perception has become a hot research object in academia,and acoustic event detection is one of the core technologies of non-speech perception.Acoustic event detection is a process of labeling temporal regions within a test audio recording and resulting in a symbolic description such that each annotation gives the timestamps and sound event labels.Due to the large variation of time-frequency characteristics in various sound events and sound fields,the problem of static background noise and overlapping of sound events has been highly valued by the research community and many evaluation activities have been carried out.Faced with the above problems,this paper has carried out research on the method of acoustic event detection.In order to realize the detection of acoustic events and improve the accuracy of acoustic event detection,this paper uses MLP and GMM classifiers to classify and detect six acoustic events,and obtains that the average event F-score of MLP is 4.6% higher than that of GMM classifier.The error rate is 2% lower than that of GMM.The correlation between the two classifiers is compared by the experimental results,the shortcomings of MLP and GMM in the detection of acoustic events are analyzed,it is also proved that the neural network based on supervised learning is superior to the non-supervised learning GMM clustering method in the acoustic event detection task of fixed data sets.In order to further improve and improve the acoustic event detection method,this paper proposes a classification method based on CRNN-HMM for acoustic event detection.In the feature extraction stage,the MFSC features are extracted for audio.In the construction stage of the acoustic classifier,the convolutional cyclic neural network is applied to the frequency dimension of the speech signal for acoustic modeling.The hidden Markov model is used to process the time dimension.The relationship is applied,and the long-term context-dependent state of the convolutional cyclic neural network is applied to process the correlation sequence between adjacent speech frames,that is,the combination of the three methods of CNN,RNN and HMM to detect and classify the acoustic events.In this paper,GLU is used as the activation function in CNN,and GLU is used in the audio classification to introduce the Attention mechanism into all layers of the neural network.The GLUs can control the flow of information in neural networks.In this way,Audio events and ignore irrelevant sounds.Experimental results of mixed audio data based on background audio and different sound events of different scenes show that when using the CRNN-HMM model,the average event F score is 7.97% higher than that of the baseline system,compared with the Bi-LSTM method.The average F score increased by 8.17% and the error rate decreased by 20%.Compared with the hybrid model DNN-HMM,the error rate increased by 4% and the F score increased by 3.67%.