Robust Speech Recognition

The sub-group on robust speech recognition specialises in investigating approaches to improve the performance of automatic speech recognition systems when deployed in hostile environments. Hostile environments are defined as environments which cannot be predicted in advance (i.e. will always be mismatched with the training data) due the presence of unknown additive noise, interfering speakers and reverberations. Such hostile environments include practical application areas of automatic speech recognition such as: dictation in noisy offices, directory assistance from roadside kiosks, call telephony from mobile calls in a social environment (café, pub, club), meeting rooms, hands-free driving assistance, etc. The methodologies being investigated included: extraction of features robust to different types of noise, adaptation of model parameters to additive and reverberant noise environments, multi-channel (microphone array) separation of speakers from interfering speakers and noise using either direct enhancement of the speech or missing data approaches.