Real-time on-board transcription of live audio announcements in trains

越来越多的乘客不听火车上的广播。乘客可能会因为听力问题、分心、背景噪音等原因错过广播。作为乘客信息系统供应商，Televic GSP旨在为** **乘客提供最新、最正确的信息。为了将有听力问题的乘客、分心的乘客或非母语乘客也包括在内，Televic GSP目前对实时音频广播字幕的研究在乘客信息屏幕上提供了可视的音频公共广播。这种离线的车载解决方案将列车工作人员发出的广播转录下来，并将其实时字幕显示在分布在整个铁路车辆的乘客信息屏幕上。2023年3月，在曼彻斯特机场至爱丁堡的列车上，与transennine Express合作进行了初步可行性研究。然而，为了进一步提高实时字幕的准确性，还需要进行更多的研究。本项目研究如何改进语音转文本解决方案的初始版本，使转录达到乘客信息系统可接受的质量。目前最先进的语音到文本模型接近90%到95%的单词准确率。然而，这种准确性是在一般口语文本中实现的。在抄写铁路公告时，准确率显著下降，特别是对于车站名称或品牌特定词等关键词：例如Hitchin vs. kitchen；运送鲸鱼vs.运送威尔士…在这个研究项目中，Televic GSP将研究离线轻量级自然语言AI模型，以进一步提高转录精度，特别是对于电台名称，警报和品牌等关键词。Televic GSP专门从事列车上的信息系统，并专注于两种类型的通信：乘客信息系统和列车监控、控制和安全。乘客信息系统部设计和维护系统，为乘客提供可靠、实时和相关的信息。Televic GSP为大型列车制造商和列车运营公司开发垂直集成解决方案，将硬件、软件和服务相结合。这些解决方案处理从路边信息管理到乘客和司机在火车上的沟通。Televic GSP在铁路行业拥有超过35年的经验，全球超过60,000辆车辆使用Televic GSP的车载乘客信息技术。在英国，这包括火车运营公司，如：Transport for Whales， Northern Trains, transennine Express Limited, Govia Thameslink Railway, Eurostar…

Increasingly more passengers don't capture audio announcements on trains. Passengers might have missed the announcement due to hearing problems, distractions, background noise, ... As a Passenger Information System provider, Televic GSP aims to provide **all** passengers with the latest, most correct information possible. To also include passengers with hearing problems, distracted passengers or non-native speaking passengers, the current research into live audio announcement captioning by Televic GSP provides the auditive public announcements visually on passenger information screens. The off-line, on-board solution transcribes the announcement made by train personnel and visualizes it as real-time subtitles on passenger information screens distributed throughout the railway vehicles. An initial feasibility study was demonstrated in collaboration with TransPennine Express on board the Manchester Airport - Edinburgh train in March 2023\. However, to further improve the accuracy of the real-time subtitling more research is still required.This project researches ways to improve the initial version of the speech-to-text solution to bring the transcription to an acceptable quality for passenger information systems. Current State-of-the-art speech-to-text models near 90% to 95% word accuracy. However, this accuracy is achieved on general spoken text. When transcribing railway public announcements the accuracy drops significantly, specifically for key words such as station names or branding specific words: e.g. Hitchin vs. kitchen; Transport for Whales vs. transport for wales, ... During this research project Televic GSP will research off-line light weight natural language AI models to further improve the transcription accuracy, specifically for key-words such as station names, alerts and branding.Televic GSP specializes in on-board information systems for trains and focusses on two types of communication: passenger information systems and train monitoring, control and safety. The passenger information systems division designs and maintains systems for reliable, real-time and relevant information to passengers. Televic GSP develops vertically integrated solutions that combine hardware, software and services to large train builders and train operating companies. These solutions handle everything from wayside information management down to the passenger and driver communication on the train. Televic GSP has more than 35 years of experience in the rail industry and more than 60 000 vehicles worldwide use on-board passenger information technology from Televic GSP. In the UK this includes train operating companies such as: Transport for Whales, Northern Trains, TransPennine Express Limited, Govia Thameslink Railway, Eurostar, ...