基于卷积神经网络与时延神经网络加权特征融合的语种识别OA

Targeting at the low recognition rate of traditional Convolutional Neural Network(CNN)and Time-Delay Neural Network(TDNN)network,a new language recognition network model is proposed.To avoid the phenomenon of neuron death during training.The frequency domain attention mechanism is introduced on the basis of the traditional CNN network and in the output of the convolution layer,the value on the negative axis of the ReLU activation function is taken by an absolute value and multiplied by a minimum con-stant.Then,on the basis of ECAPA-TDNN network,a pooling layer is introduced after the first layer of convolution to remove redundant information in features.Finally,weighted fusion is performed between the constructed FCA-CNN network and NECAPA-TDNN network to form a fusion network,and the extracted speech features are used as the input of the fusion network for classification and recognition.The experimental results show that the average recognition accuracy of the spectrogram features in LibriVox data set is improved by 10.96%and 17.12%compared with CNN and TDNN,which verifies the effectiveness and recognition performance of the fusion network.