Open speech resources application in sports distance courses based on recurrent neural network

Speech recognition in low-resource training data has gradually become a key research problem in general speech recognition under limited conditions in recent years because of its low recognition rate. The original state has a great influence on the two modules of feature extraction and acoustic model. How to effectively improve the training effect of the deep learning model and how to improve the deep learning model are key issues to be solved urgently. This paper uses deep learning technology to research and improve the characteristic information and acoustic model in the speech recognition system created by the author. In this article, we propose a method to improve the author's convolutional neural network acoustic model by fusing multi-stream functions. In a speech recognition system with sufficient training results, the DNN acoustic model has a lower recognition rate than the CNN acoustic model. However, in the absence of resources and training data, there is insufficient training of network parameters. To use more acoustic feature information to model from limited data, first extract several types of features from resource-poor training data, then construct parallel convolutional subnets with different feature types, and finally pass them through completely the connected layers are merged to create a new CNN acquisition structure.