Abstract: In order to standardize the verbal instructions of railway dispatchers, this paper planned to use keyword recognition technology to identify and record the operation instructions of dispatchers, and restrict and adjust the non-standard terms. Based on the recognition technology of keyword-waste model, the paper used the end-to-end training method of maximum mutual information criterion without word lattice dependence to train the keyword recognition model, so as to realize the automatic recognition of railway dispatcher's oral instructions. Experimental analysis was carried out for the two keywords with the highest frequency to study the feasibility of this method in industrial scenarios. The experimental results show that by adjusting parameters to control the error recognition rate, a certain keyword recognition rate can be achieved which is feasible for application.