Boosting of Contextual Information in ASR for Air-Traffic Call-Sign Recognition

Contextual adaptation of ASR can be very beneficial for multiaccentand often noisy Air-Traffic Control (ATC) speech. Ourfocus is call-sign recognition, which can be used to track conversationsof ATC operators with individual airplanes. Wedeveloped a two-stage boosting strategy, consisting of HCLGboosting and Lattice boosting. Both are implemented as WFSTcompositions and the contextual information is specific to eachutterance. In HCLG boosting we give score discounts to individualwords, while in Lattice boosting the score discountsare given to word sequences. The context data have origin insurveillance database of OpenSky Network. From this, we obtainlists of call-signs that are made more likely to appear inthe best hypothesis of ASR. This also improves the accuracyof the NLU module that recognizes the call-signs from the besthypothesis of ASR.As part of ATCO2 project, we collected liveatc test set2.The boosting of call-signs leads to 4.7% absolute WER improvementand 27.1% absolute increase of Call-Sign recognitionAccuracy (CSA). Our best result of 82.9% CSA is quitegood, given that the data is noisy, and WER 28.4% is relativelyhigh. We believe there is still room for improvement.