Design And Implementation Of Music On Demand System

After the explosive growth in recent years,intelligent speakers,as the hottest AI product at present,have been outstanding in hardware products for a long time.For the speaker,the most important function module is the music on demand system.This paper introduces the music on demand system in detail through four processes:analysis,design,implementation and testing.During my internship in the company,I personally participated in the demand analysis phase,system design phase,implementation phase and testing phase of the project.I am mainly responsible for the design and specific development of the system and self-test after development in the project.Through the analysis of the use of the previous system,we can see that its main problems are as follows:(1)the accuracy and real-time performance of the system for users' speech intention recognition is not up to the basic requirements;(2)the lack of music slot value vocabulary is not conducive to the training of music-related models,resulting in low recognition rate;(3)Because of the limitation of service capability of service providers,the whole business chain process is too long,which affects the overall response time of services.The final needs of the project to solve the above three problems are obtained.Aiming at the three main problems of the old system,the first problem is solved by designing a new language model.By crawling a large number of network music resources and building an efficient Redis database,we can solve the problem of no slots word-list.By designing ES and Redis databases with high reading efficiency,the problem of high total latency of the system is solved.After the completion of the design,I am responsible for the specific implementation of the project,mainly responsible for all the development of the system,including the construction of the database,the construction of the new language model and the construction of the system process,and test and improve.The innovation of the project is to train Bi-LSTM-CNN-CRF joint model with intention recognition and slot extraction tasks,so as to get better result sequence.At present,the system has been completed and operated well.With the help of this system,users have the experience of playing songs on demand through non-modal language at will.