Research And Implementation On Constructing Speech Collection System Based On Deep Learning

With the development of disciplines such as computer science and communication,speech recognition technology has gradually grown and is becoming more and more mature.Especially in recent years,the re-emergence of machine learning methods has made speech recognition technology a further breakthrough.Nowadays,speech recognition technology has been widely used in various fields,such as mobile phone voice assistant,voice housekeeper of smart home,robot voice.Customer service,etc.In practical applications,different scenarios have different emphasis on the performance requirements of speech recognition technology.Among them,the robot voice customer service is in the call scene,which requires the robot customer service to respond quickly when understanding the user's intention.However,the relevant performance of the voice recognition technology in the field of robot customer service needs to be improved.Therefore,this paper proposes a lightweight deep learning-based speech matching model.Different from the traditional speech recognition model,the speech signal is converted into text characters.This model directly inputs the feature vector obtained by the feature extraction algorithm into the deep neural network.In the model,the similarity between the feature vectors is compared,and the matching result is output.And in the application phase,a robot voice customer service system for collection business was built.The specific work of this thesis includes:(1)In-depth study on speech endpoint detection algorithm,speech feature extraction algorithm and deep learning algorithm,through experiment and comparative analysis of the advantages and disadvantages of similar algorithms,and make a choice.(2)Through a large amount of literature reading and practice,the current speech recognition model framework is analyzed and summarized,and the speech matching model for specific task scenarios is studied and designed.The deep learning-based speech matching model designed in this paper is used in specific tasks.In the scenario,the response speed and accuracy are better than other models;(3)Based on the speech matching model and other speech recognition related technologies designed in this paper,a robot customer service system for collection business is built.The data of this subject comes from Merchants Union Consumer Finance Company Limited.Through the experimental comparison,the voice matching model proposed in this paper to speech recognition systems on the market is batter in both the recognition accuracy and the response time in the relatively fixed business scenarios.