Large Vocabulary Speech Recognition
The task of large vocabulary continuous speech recognition requires advances to be made in several areas of spoken language processing. Our research in this area is based around a hybrid neural network/hidden Markov model acoustic modelling system. In a collaboration with Cambridge University, the resultant system has been one of only two UK systems to take part in the ARPA/NIST evaluation of speech recognition systems. The hybrid approach offers both modelling and computational advantages compared with the more usual HMM approaches. Active research areas include efficient search, domain adaptation, statistical language modelling, speaker adaptation and improved neural network probability estimation approaches.