Towards Continuous Speech Recognition for BCI

Abstract

For the last two decades, brain-computer interface (BCI) research has worked towards practical and useful applications for communication and control. Yet, many BCI communication approaches suffer from unnatural interaction or time-consuming user training. As continuous speech provides a very natural communication approach, it has been a long standing question whether it is possible to develop BCIs that perform speech recognition from cortical activity. Imagined speech as a BCI paradigm for locked-in patients would mean a large improvement in communication speed and usability without the need for cumbersome spelling using individual letters. We showed for the first time that automatic speech recognition from neural signals is possible. Here, we evaluate the feasibility of speech recognition from neural signals using only temporal offsets associated with speech production and omitting information from speech perception. This analysis provides first insights into the potential usage of imagined speech processes for speech recognition, for which no perceptive activity is present.