语音识别中的DenseNet模型研究

为了解决语音识别中由网络加深导致的低层特征消失、参数量大及网络训练困难的问题，基于Inception V3网络的非对称卷积思想，提出了一种改进的密集连接卷积神经网络(densely connected convolutional neural networks, DenseNet)模型。根据语音识别的长时相关性，通过密集连接块建立起不同层之间的连接关系，从而保存低层特征、加强特征传播;为了得到尺度更丰富的声学特征，将卷积核的范围进行扩大;利用非对称卷积思想分解卷积核，以减少参数量。实验结果表明，相较经典深度残差卷积神经网络模型和原始DenseNet模型，提出的模型在THCHS30数据集上的语音识别性能更好，在保证识别率的情况下，还减少了网络参数量，提高了模型训练效率。

Research on the DenseNet model for speech recognition

To solve the problems of the disappearance of low-level features, the large amounts of parameters and the difficulty of network training due to the deepening of the network in speech recognition, we propose an improved densely connected convolutional neural network model (DenseNet) based on the idea of asymmetric convolution of Inception V3 network. According to the long-term correlation of speech recognition, the model uses dense connection blocks to establish connection relationships between different layers to preserve low-level features and strengthen feature propagation, and in order to obtain a richer scale of acoustic features, the model expands the scope of the convolution kernel. In addition, the idea of asymmetric convolution is used to decompose the convolution kernel to reduce the amounts of parameters. The experimental results reveal that compared with the classic deep residual convolutional neural network model and the original DenseNet model, this model has better speech recognition performance on the THCHS30 data set. In the case of ensuring the recognition rate, the amount of network parameters is reduced, and the training efficiency of the model is improved.