bioRxiv2022
BackgroundBrain-computer interfaces (BCIs) can offer solutions to communicative impairments induced by conditions such as locked-in syndrome. While covert speech-based BCIs have garnered interest, a major issue facing their clinical translation is the collection of sufficient volumes of high signal-to-noise ratio (SNR) examples of covert speech signals which can typically induce fatigue in users. Fortuitously, investigations into the linkage between covert speech and speech perception have revealed spatiotemporal similarities suggestive of shared encoding mechanisms. Here, we sought to demonstrate that an electroencephalographic cross-condition machine learning model of speech perception and covert speech can successfully decode neural speech patterns during online BCI scenarios. MethodsIn the current study, ten participants underwent a dyadic protocol whereby participants perceived the audio of a randomly chosen word and then subsequently mentally rehearsed it. Eight words were used during the offline sessions and subsequently narrowed down to three classes for the online session (two words, rest). The modelling was achieved by estimating a functional mapping derived from speech perception and covert speech signals of the same speech token (features were extracted via a Riemannian approach). ResultsWhile most covert speech BCIs deal with binary and offline classifications, we report an average ternary and online BCI accuracy of 75.3% (60% chance-level), reaching up to 93% in select participants. Moreover, we found that perception-covert modelling effectively enhanced the SNR of covert speech signals correlatively to their high-frequency correspondences. ConclusionsThese findings may pave the way to efficient and more user-friendly data collection for passively training such BCIs. Future iterations of this BCI can lead to a combination of audiobooks and unsupervised learning to train a non-trivial vocabulary that can support proto-naturalistic communication. Significance StatementCovert speech brain-computer interfaces (BCIs) provide new communication channels. However, these BCIs face practical challenges in collecting large volumes of high-quality covert speech data which can both induce fatigue and degrade BCI performance. This study leverages the reported spatiotemporal correspondences between covert speech and speech perception by deriving a functional mapping between them. While multiclass and online covert speech classification has previously been challenging, this study reports an average ternary and online classification accuracy of 75.3%, reaching up to 93% for select participants. Moreover, the current modelling approach augmented the signal-to-noise ratio of covert speech signals correlatively to their gamma-band correspondences. The proposed approach may pave the way toward a more efficient and user-friendly method of training covert speech BCIs.