Research On Railway Dispatching Speech Recognition System Based On BiLSTM

The safe operation of railways is of great significance to promoting economic growth,improving people’s livelihood and strengthening national defense security.Railway dispatching is the "central nerve" of railway operation,and the dispatching telephone between the train dispatchers and the train driver,signal and communication department,engineering department and other railway operation and production units is the core part of railway operation organization process.Previous speech recognition research mainly focuses on the recognition of common language and everyday speech,while the application research in the field of railway focuses on the recognition of non-secure speech content such as guiding ticket purchase and passenger transport service.However,the research on railway dispatching speech recognition which directly affects traffic safety and efficiency is seldom discussed.This study takes railway dispatching speech as the research object and carries out an in-depth study on railway dispatching speech recognition in the training and assessment process of railway traffic virtual simulation experimental platform.This study enriches the application research of speech recognition in the railway field,and provides a more comprehensive evaluation index for the training and assessment of railway dispatching personnel,which has certain practical significance and practical value.Firstly,combining with the working principle of the existing speech recognition technology,the types and system framework of railway dispatching speech recognition are further defined and designed.In order to solve the problems of mixing environmental noise and redundant information in the process of speech signal production,transmission and collection,the original speech signal is processed and analyzed by preprocessing and feature extraction,which will get FBank feature and MFCC feature used in railway dispatching speech recognition system input,it lays a foundation for the construction of acoustic model of railway dispatching.In order to simulate the communication content between front-line vehicle personnel,a combined voice data set based on railway dispatching communication data set was constructed according to the standard library of station-locomotive joint control terms.According to the application environment of railway dispatching terms,the applicability of different language models is analyzed and selected.In order to solve the problem of the particularity of railway dispatching speech in pronunciation and intonation and the different importance of words in dispatching terms,a corresponding solution to dictionary annotation is proposed.Secondly,in order to improve the accuracy of railway dispatching speech recognition,appropriate acoustic models are built.The Gaussian mixture model(GMM)and the deep neural network(DNN)were used to construct the acoustic model of railway dispatching based on the Hidden Markov Model(HMM),respectively.GMM/DNN was used to fit observation probability distribution and other probability parameters of HMM were used as training objects.An acoustic model of railway scheduling based on BiLSTM was constructed by combining BiLSTM with CTC.The BiLSTM was trained with CTC as the loss function.Through the comparison of word error rates,it is proved that the acoustic model based on BILSTM-CTC has outstanding advantages in the field of railway dispatching speech recognition.Finally,in order to solve the "dispatching speech-operation" consistency check requirements of the railway traffic virtual simulation experimental platform,the recognition text is post-processed with the railway dispatching speech recognition results as the input.The recognition text post-processing is carried out with the railway dispatching speech recognition results as input.According to the keywords and location characteristics of railway dispatching terms,the corresponding semantic solution of railway dispatching is proposed,which can avoid the misjudgment of the identification results of the experimental platform to a certain extent and improve the intelligent level of the experimental platform.