Use of Linear Descriminant Analysis in Isolated and Continuous Speech Recognition for an AIr Traffic Control Task

This paper describes experiments with a speech recogniser using whole-word models and linear discriminaan for its spectral representation in the framework of an Alvey demonsu-ator for an air traf c control task.The demonstrator task and speci cation are rst set out.After a brief account of the speech output componcnl.the implementation of the recognition hardware is described.Three series of recognition tests are then described.In the rst, the performance of the new hardware is shown to be superior to the best previous hardware in speaker independent tests with normal and with speech in which the loudness and voice quality is varied.Inthe second.the absolute performance of the system is tested on sentences from the demonstrator task.Finally. the third series provides results to allow comparison on a common database with a sub-word modelling system.