Recognition of Oral Speech from MEG Data Using Covariance Filters

Recognition of spoken speech based on EEG and MEG data is the first step in the development of brain– computer interface (BCI) and artificial intelligence (AI) systems for their application to decoding imagined speech. Major advances in this direction have been made using ECoG and stereo-EEG techniques. However, there are few studies on this topic using analysis of data obtained by non-invasive methods of recording cerebral activity. The approach presented here is based on assessment of connections in sensor space with identification of MEG connectivity patterns specific for particular segments of speech. The method was tested in seven subjects. The processing pipeline was quite robust in all cases, running with no or few recognition errors. After training, the algorithm was able to recognize a fragment of oral speech with just one presentation. Recognition used MEG recording segments 50–1200 msec from the onset of the sound of the word. High-quality recognition required a period of at least 600 msec. Intervals longer than 1200 msec worsened recognition quality. Band-pass filtering of MEG showed that recognition quality was equally effective over the entire frequency range. A slight decrease in recognition level as seen only in the range 9–14 Hz.