Study On Mask Speech Recognition Method Based On Neural Network

Full face mask is one of the necessary equipments that divers use underwater.With the thoroughly exploring of the ocean,more and more intelligent devices are used in shallow water acoustic communication and operations.And mask speech recognition is used as an important interface for human-computer interaction,which brings great convenience to the divers.Firstly,mask speech recognition is of great significance to improve divers’ efficiency of underwater work.What’s more,it can also ensure divers’ life safety.Considering the actual environment,there will be variety of ocean noises,which affect the normal conduct of human-computer interaction.In this thesis,the way combining speech enhancement algorithm and speech recognition algorithm is used to recognize the noisy mask speech.Firstly,enhancement algorithm is used to enhance the noisy mask speech.Then,recognition algorithm is used to recognize the mask speech that enhanced.The main work and research contents are as follows:(1)Speech signal preprocessing.Firstly,by comparing the spectrogram of mask speech and air speech to analyze the characteristics of mask speech,the result shows that the high frequency component of mask speech is weaken or even lost.Then preprocess the mask speech and extract the feature parameters.(2)Research on mask speech enhancement algorithm.An improved wiener filtering algorithm is proposed in this thesis.The spectrum entropy algorithm is used to detect the state of each frame of the mask speech before calculating the gain function,and update the noise power spectrum for the frame without speech.At the same time,the gain control parameter is introduced to reduce the mask speech distortion when enhancing the mask speech.By comparing and analyzing the time domain waveform and spectrogram of enhanced mask speech,the result shows that the mask speech that enhanced by the improved wiener filtering algorithm has less residual noise when the gain control parameter increases.However,if the parameter is too large,the mask speech with noise will be distorted more seriously.Compared with two classical enhancement algorithms,the improved wiener filtering algorithm after choosing the reasonable gain control parameter has better performance.(3)Mask speech recognition based on neural networks.Conduct simulation experiment on the mask speech recognition by Back Propagation(BP)neural network,Convolutional Neural Network(CNN)and Long Short Term Memory(LSTM)network.And the influenceof three neural networks with different structures on the mask speech recognition rate is studied.At the same time,Local Response Normalization(LRN)and Dropout are used to optimize CNN.The result shows that the recognition rate of isolated word mask speech is higher by using CNN than using BP or LSTM.(4)Research on the method of noisy mask speech recognition.For the interference of ocean noises,firstly enhance the mask speech,then recognize the mask speech by CNN in this thesis.The structure and parameters of the model in this thesis are determined by comparing and analyzing different structures and parameters.The result shows that compared with the way combining spectral subtraction algorithm with CNN,the way combining improved wiener filtering algorithm with CNN has higher recognition rate under wave noise and seawater noise with different SNR.Compared with the way combining wiener filtering algorithm with CNN,the way combining improved wiener filtering algorithm with CNN has higher recognition rate under wave noise with different SNR and has higher recognition rate under seawater noise with high SNR,but has lower recognition rate under seawater noise with low SNR.