Study On Wake Up Word Recognition Based On Deep Learning

Wake-up-word speech recognition technology has a wide range of applications in smart audio,smart car,service robot and smart home.Deep neural network based wake-up word recognition method has made some progress and has been applied in certain scenarios.Due to the limitations of DNN itself,it does not consider the correlation between speech signals.It makes its false wake-up rate higher,which brings a bad experience to users.Its parameter quantity is relatively large,which is often difficult to meet the requirements of practical applications.The wake-up word recognition method based on the Long Short Term Memory Network cannot learn the relationship between the time domain and the frequency domain.And it has a large amount of calculation.In order to solve the above problems,this thesis mainly studies the method of recognition of wake words based on deep learning,and the main work is as follows:(1)The wake-up word recognition method based on Simple Recurrent Unit(SRU)is studied.When calculating the reset gate and the forgetting gate,the method adopts the complete discarding method and removes the dependence on the state of the hidden layer at the previous moment.The method for recognizing the wake-up word shortens the training time of the model and reduces the parameter amount of the model.It improves the accuracy of the model and improves the wake-up rate of the model.(2)The wake-up word recognition method of convolution recurrent network based on attention mechanism is studied.This method combines Convolution Neural network(CNN)and SRU network to form a CSRU network.It utilizes CNN's ability to capture local information and utilizes SRU's time series modeling ability to consider the spatial characteristics of the spectral map.On this basis,the method of recognition of wake words based on attention mechanism is studied.It consists mainly of an encoder and an attention mechanism.Encoder uses CSRU network.The soft attention mechanism learns normalized weights from the feature representation.The probability distribution of wake-up word is generated by a linear transformation and a Softmax function.This method improves the accuracy of the model and reduces the false wake-up rate of the model.