| Prison shoulders the important responsibility of implementing law and maintaining social stability in China's judicial system.It is important to build a sound security alarm system for the security.The traditional alarm system combines with manual triggering and monitoring video analysis has some problems such as high false alarm rate,lack of real-time performance and monitoring blind spots.Therefore,developing a police system based on speech keyword detection has great significance for improving the supervisory security alarm system of the prison,which can solve the above problems.In this thesis,the software of alarm system based on speech keyword detection is designed and developed for the reliability,real-time and integrity of the alarm system.The software adopts B/S architecture design,which mainly realizes the voice alarm analysis,real-time alarm pushing,and historical alarm backtracking.The software collects voice data through the device access layer,and breaks sentence through the voice activity detection based on GMM.The concurrent programming technology is used to realize the multi-channel parallel CNN-based speech keyword detection.Generate alarm or early warning events by analyzing the test results,and perform real-time push and record storage.The system improves the construction method of the training data set and the data model for the characteristics of the scene of prison.The Kaldi framework-based speech keyword detection process is optimized for the real-time alarm requirement,which improves the accuracy and execution efficiency.In addition,the system designs a streaming audio playback framework based on MSE+WebSocket,to meet the needs for real-time voice playback and alarm record playback.After testing,the system's speech keyword detection accuracy rate is over 97%,the 8-way parallel voice alarm analysis task takes 1.2s?1.6s,and the real-time audio delay is less than 1.5s,which satisfies the system in function and performance demand. |
PDF Full Text Request