Developments in speech recognition for ATC

Speech recognition has been successfully used at DFS (German Air Navigation Services) in training of Air Traffic Controllers since 2010. An operational application promises significant benefits but simultaneously requires high reliability in speech recognition. The detection of specific keywords already works reliably in many cases. Representative operational speech data are necessary for further developing speech recognition, but they face strict access restrictions in Germany due to legal requirements (e.g. GDPR). Current DFS research and development efforts aim to address the complex situation through preprocessing of ATC speech data to remove sensitive information. The prerequisite is that high speech recognition performance is maintained for this data. The study presents the R&D concept “SALT” for using speech recognition as an input assistance for Air Traffic Controllers at DFS to reduce workload. Moreover, it shows the results of an initial speaker identification experiment revealing that anonymization filters can reduce the hit rate down to 20%. However, chance level for speaker identification was not yet reached. More testing is needed to be able to demonstrate that the procedure is safe enough and to achieve access to operational speech data for training of speech recognition algorithms.