Transforming Imagined Thoughts into Speech Using a Covariance-Based Subset Selection Method

Abstract

With the advancement of engineering solutions in the medical domain, the patient’s life can become comfortable. This work recognizes the silent speech of three words. The decoding of silent speech can be useful for patients who are in a locked-in syndrome state. Moreover, it is also applicable to entertainment, cognitive biometrics, and brain-computer interfacing. Brain waves of these imagined words in the delta, theta, alpha, beta, gamma, and high gamma frequency bands are analysed. Covariance based connectivity features are extracted in each frequency band. The principal features which represent more than 95% of the variance are selected as a subset of the covariance connectivity matrix. This sub-set is tested on five classifiers. The maximum accuracy achieved is 76.4% in the theta band. Also, theta and high gamma band contain maximum information about imagined speech with average accuracies of 68.32% and 62.09% respectively.