Optimization And Design Of Speech Recognition Scheme Based On Recurrent Neural Network

Automatic speech recognition is an important branch of artificial intelligence.As a signification human-computer interaction technology,speech recognition has been widely used in most kinds of life scenes in recent years.The accuracy of the speech recognition system in pure audio environment has exceeded people's hearing.However,the complex environment factors in real application scenarios affect the speech recognition system most.And the time spending and the computational costing in development stage hinder the advance of speech recognition technology seriously.Therefore,improving the accuracy of speech recognition system and reducing the time spending of model training simultaneously is an important topic for speech recognition.The non-interpretability of the neural network is the main obstacle of the neural network optimization.In this paper,we employ the visualization method to exploring the internal structure of the neural network.And the primary reason of the low performance and slow training speed are analyzed based on the visualization results.The research work of this paper consists of three parts.In the first part,a deep neural network based on cross layer value transferring is proposed.The increasing of the hidden layer's quantity eliminates over fitting effectively.However,the information losing in the process of transmission led by it.According to the study of the hidden layer this paper deliver a new deep structure based on cross layer value transferring.It is proved that such method can prevent the information losing through the process of transmission effectively.In the second part,a method based on binarization and linear representation is proposed to speed up the training and decoding process.The complex structure of neural network is the primary cause which lead to the slow computational of the neural network.We proposed a method based on combine the binarization and linear representation strategy,which is proved that such method can accelerate the training and decoding process with a tiny accuracy lost.In the third part,we combine the above two methods and propose a new binarization and linear representation recurrent neural network architecture based on cross layer value transmission,after which three test methods are carried out which are conventional test,robust test and speed test.The results show that the new model can enhance the performance of speech recognition in the premise of accelerating the training and decoding process.Finally,a simple online speech recognition system is implemented base on the new recurrent neural network architecture.Such system can provide online speech recognition service for users.