Consistency Modeling



The main goals of this project are to develop new methods to train speech recognition models so as to significantly improve accuracy without the need for large amounts of data, and new ways to represent models and share parameters so as to significantly increase recognition speed and decrease model size while maintaining accuracy. Another important goal is to develop methods to quickly port speech recognition systems to domains with different environmental conditions than the training domain, such as different channel and noise conditions, and non-native speech. Finally, the project is concerned with rudimentary information extraction from the recognized speech by automatically identifying information containing entities in the speech, such as person, organization and place names.