Analysis of cortical activity for the development of brain-computer interfaces for speech

Overview & Synopsis

Brain-computer interfaces for speech have the potential to restore the ability to speak in people with severe paralysis. Such a system would decode features of what the user wants to say from their brain activity and synthesize the associated sounds in real time. Recent studies show promising results in this direction by decoding speech from brain recordings of healthy participants speaking aloud. However, several challenges must be overcome to enable speech decoding in people unable to speak. This thesis aims to contribute to the development of brain-computer interfaces and to the understanding of cortical activity underlying speech.First, we highlighted a phenomenon of acoustic contamination of electrophysiological signals during speech. Our analyses revealed that this phenomenon was present in several recordings from different laboratories around the world and that it could have an important impact on speech decoding studies by biasing the obtained performances. We studied the origin of this phenomenon and proposed a methodology to quantify it.It is expected that the cortical activity of paralyzed subjects attempting to speak is partially different from that of healthy subjects speaking aloud, in particular because of the absence of auditory and somatosensory feedback. Understanding the differences between these two conditions is important to know if the models currently developed on overt speech would be easily applicable to paralyzed people. In this perspective, studying covert speech in healthy participants allows us to study the activity underlying speech production in the absence of sensory feedback, as it is the case for people with paralysis.Using an intracortical recording, we showed that the dynamics of the activity of a population of neurons in Broca's area were different during overt and covert speech. Moreover, we found that the temporal arrangement of action potentials recorded during the pronunciation of sentences allowed to discriminate these sentences. The similarity of spike trains appeared to be correlated with the phonological similarity of the sentences. Finally, using the same recording, we showed that the low frequency component of local field potentials was modulated by overt speech and to a lesser extent by covert speech.Another obstacle concerning the decoding of attempted speech in participants unable to speak is the difficulty to calibrate the decoder. Indeed, in the absence of behavioral output, it is not possible to know precisely the timing at which the participant attempts to speak. We therefore tested, in a non-paralyzed participant, a protocol allowing to build a covert speech decoder from electrophysiological activity. Our offline analyses showed that it was possible to decode acoustic features of imagined phonemes with performances statistically superior to chance. We also performed a preliminary test of real-time, closed-loop decoding in the same participant.