A survey on EEG-based imagined speech classification

Allowing communication in those situations in which the use of the voice or other human expressive means is not possible is one of the main objectives of two broad research areas: silent speech interfaces and brain computer interfaces. They are mainly focused on disabled persons, but they are also of interest for healthy people. The intersection of these areas lies in the aim of recognizing imagined speech recorded using electroencephalograms. For BCIs, imagined speech offers a neuroparadigm more intuitive and with more degrees of freedom than the most used BCIs. In this study, we survey all works in this area and it is different to others in that it is only focused on the works that look for recognizing imagined speech from EEG signals. Like common speech processing theories, these works have approached this task following two broad paths: short vocalizations (syllables, phonemes, and vowels) and words. This interesting problem remains open, although several initial steps have been taken as to the application of novel techniques from pre-processing, artifact removal, feature extraction, and classification. Also, the community recently have shared datasets looking for defining a gold standard for comparison purposes. Finally, despite most of the works having studied the problem in an offline mode, a few works have started to face the online imagined speech recognition problem.