Incorporation of a language model into a brain-computer interface-based speller through HMMs

Abstract

Brain computer interfaces (BCI) is a well-known application of human computer interaction that has been designed as an assistive technology for people who are not able to use any motor system for communication. The P300 speller is a widely used example of a BCI typing system utilizing daily language. Due to the low signal-to-noise ratio (SNR) in electroencephalography (EEG), several numbers of trial groups are needed to improve accuracy, which also results in low speed of the system. To this end, we have proposed to construct a hidden Markov model (HMM) for the integration of a trigram language model into EEG classification scores obtained from Bayesian Linear Discriminant Analysis. With integration of a language model, the results indicate that our model can increase the accuracy and bit rate up to %35 and %55, respectively, in the first five trial groups. We also added Gaussian noise to the EEG data to observe the efficiency of the language model in poor conditions and experimental results on this issue exhibit the improvements in performance values.