Speech-Recognition Software for Air Traffic Control

Air traffic control (ATC) is a complex task that is taxing on the human operator. It has long been recognized that voice input can greatly simplify the operator's task and reduce fatigue. Current real-time ATC training systems rely on expensive, special-purpose hardware. A portable all-software solution would reduce costs and be more readily modified for future needs. In Phase I, we propose to develop a baseline continuous-speech system using our novel single- and multi-stream approach. We have previously demonstrated that this approach lowers error rates up to twenty-fold for isolated-word recognition, in particular when noise is intense and differs between training and testing-a traditionally difficult problem. We will compute the trade-off curves between accuracy, speed of computation, the size of the vocabulary and grammar perplexity. By trading off the increased accuracy for speed, we predict real-time performance for an all-software solution. We will also apply the multi-stream approach to the speaker-independent domain. In Phase I, we will build a continuous-speech system based on the single/multi-stream approach. The work will provide a foundation for Phase II, where we will extend the baseline system to the full ATC vocabulary and deploy and test prototypes.