Research On Triggering Method Of Taxi Emergency Call System Based On Speech Recognition

Nowadays, the taxi-related traffic accidents and robberies occur frequently,which has not only caused wide concern by taxi drivers but also highly emphasized bypublic security departments. To shorten the rescue time, reduce casualties, timelycapture criminals, and safeguard the taxi driver’s life and property, a new taxiemergency call system is proposed. It can automatically call the rescue center for helpafter road traffic accidents or looting. At the same time, the danger alarm flashlight oftaxi is open, and the location information with danger mark is uploaded in real-time.To make up for shortcomings of existing triggering methods, a new triggeringmethod based on speech recognition is proposed. It can use speech recognitiontechnology to determine whether the current scene is dangerous by recognizing callfor help and robbery. If it is identified as a dangerous scene, the system willautomatically call the rescue center for help. Once get the call, rescue center listens tothe voice inside the taxi, and gets images of the taxi. After re-recognition, rescuecenter sends rescue agencies to help. This triggering method is not easy to be foundby robbers, more importantly, it can safeguard taxi driver’s life and property, which isof great importance on application value and social development.When studying the triggering method based on speech recognition, the followingwork are done:(1) The characteristics of call for help and robbery scene are analyzed. The basicprocess of triggering method based on speech recognition is designed. When akeyword, such as “help me” or “don’t kill me”, etc., is recognized by the speechrecognition module, the keyword sequence in recent a period of time is trimmed. Thenthe dangerous scene recognition algorithm is invoked to determine whether thecurrent scene is dangerous. If it is, the alarm module is triggered.(2) The basic technologies of speech recognition and the speech enhancementalgorithm based on MMSE-LSA are studied. The frequency-domain smoothingmethod of noise estimation based on connected time-frequency speech presenceregions is modified. A speech recognition module based on PocketSphinx is designed.The performances of PocketSphinx are tested on the Android smartphone. The experimental results show that the speech enhancement module can improve therecognition rate of PocketSphinx on the condition that noise is in existence in the car.The recognition rate is93.7%.(3) A dangerous scene recognition algorithm based on statistics and fuzzymatching is designed. Its design rules of pattern library are formulated. All patternsare sorted by similar number thresholds in ascending order. This algorithm uses thenumber of keywords to determine whether the scene is dangerous. If it is, thenperforms over, otherwise traverses the pattern library. If the number of keywordsdoesn’t meet with the minimum requirements of current pattern, the current scene isn’tdangerous and traverse is over, otherwise the greedy string tiling algorithm is utilizedto calculate the similar number between keywords sequence and the pattern. If thesimilar number meets the requirements of the pattern, then the current scene isdangerous and traverse is over, otherwise continue to traverse, until all patterns aretraversed. Simulation experiments based on Android smartphone show that therecognition rate of this algorithm is94.7%, the leakage rate is5.3%.