We are developing speech technology to provide a natural interaction with information, services and technology.
Today's speech systems allow users to do many things from searching the web on their smart-phones to listening to e-books as they travel. However, there are still many challenges to solve before the full power of speech communication is achieved.
The advent of statistical parametric speech synthesis has opened up the possibility of creating rich, expressive synthesised speech. This will allow us to convey information more efficiently, as well as making the speech far more interesting and pleasurable to listen to. At CRL we are at the forefront of research in this area.
We have pioneered the application of acoustic factorisation in speech synthesis with a system that allows the speaker and language to be controlled separately. A user can "speak" in several languages for personalised voice translation.
In the domain of expressive synthesis, we have developed Xpressive TalkTM, a digital talking head which can express human emotions, in collaboration with the Computer Vision Group and the University of Cambridge's Department of Engineering.
The two factors that most affect speech recognition performance are the huge variations across speakers, and background environmental noises. In collaboration with the University of Cambridge, we have invented world leading techniques to make recognition systems more robust by jointly adapting to the speaker and noise condition.
Our technology has been applied to an exemplar system which records and transcribes our meetings. We have created an integrated approach to speaker adaptation and diarization to provide high quality transcriptions of what was spoken, by whom and when. Our multi-national team means the system has to handle many different native and non-native accents of English. In addition, it has to cope with reverberant noise from the room acoustics and background noises such as the projector and air conditioning.
The STG Publications page contains links to papers we have published on these topics.