Research On Speech Recognition System In Vehicle's Environment Of Rail Transit

The safe operation of rail transit is of great significance to China's economic development,improvement of people's livelihood and national defense security.Among them,the dialogue between the conductor and the management of the dispatching center is the most core part in the safe operation of the train.Due to the strong work intensity of the train staff and the noisy conversation environment,it is easy to have vague conversation voice,which leads to operational errors and threatens the safe operation of the train.Therefore,the voice automatic recognition technology arises at the historic moment.It can automatically detect the train attendant's various operation behaviors,not only give early warning of the danger caused by wrong instructions,but also provide convenience for investigation and evidence collection after accidents.It will be an important means to ensure the safe operation of trains in the future.However,on the one hand,because the voice of train conversation has its particularity in pronunciation,intonation and other aspects,the current common voice automatic recognition technology on the market cannot be directly used.On the other hand,the data of the train attendant's conversation voice are still difficult to collect and label,which leads to the difficulty of accurate implementation of automatic recognition technology.Therefore,this paper will carry out in-depth research on the application of automatic speech recognition technology in the field of rail transit and the improvement of recognition accuracy.(1)This paper will give a detailed introduction to the framework and principle of continuous speech recognition,including feature extraction of speech signals,speech modeling technology of speech signals,pronunciation dictionary and language model,and speech decoding network based on weighted finite state converter.(2)In order to solve the problem of unclear voice during train operation,this paper compares four different noise reduction methods with real train attendant dialogue voice signal data as experimental data.The experimental results show that the improved spectral subtraction algorithm has better noise reduction effect.(3)This paper compares the corresponding experimental results by building GMMHMM monophone and triphone models respectively.The results of this experiment show that the recognition accuracy is higher when three phonemes are used as sound modeling units.In order to eliminate the influence of environmental background noise and speaker's different accents,the maximum likelihood change and adaptive processing are carried out on the three-phoneme model,and the sound model is further improved to obtain better recognition effect.(4)Finally,through the use of LSTM neural network model to build a sound model,in the experimental part,the final recognition results of GMM-HMM and LSTM model are respectively compared,and the conclusion is: based on the train-borne environment,LSTM will have stronger modeling capability in the speech phoneme recognition process.