Abstract: Aiming at the deficiencies of conventional speech analysis technologies in railway dispatch scenarios, including the recognition of domain-specific terminology, processing of regional accents, and trade-off between real-time performance and recognition accuracy, this paper constructs a Railway Dispatch Speech Analysis Large Language Model (RDSA-LLM) by adopting customized professional hot vocabularies, regional accent adaptation schemes and real-time optimization algorithms. Relevant experiments are carried out to verify the model effectiveness. On the basis of the proposed model, practical applications covering real-time dispatch speech transcription & monitoring, intelligent emergency response assistance and compliance inspection of dispatch commands are developed. The research results can facilitate comprehensive safety management of railway dispatch and command, and provide technical support and application reference for the intelligent upgrading of railway dispatch systems.