A Communication Paradigm Using Subvocalized Speech: Translating Brain Signals into Speech

Abstract

Recent science and technology studies in neuroscience, rehabilitation, and machine learning have focused attention on the EEG-based brain–computer interface (BCI) as an exciting field of research. Though the primary goal of the BCI has been to restore communication in the severely paralyzed, BCI for speech communication has acquired recognition in a variety of non-medical fields. These fields include silent speech communication, cognitive biometrics, and synthetic telepathy, to name a few. Though potentially a very sensitive issue on various counts, it is likely to revolutionize the whole system of communication. Considering the wide range of application, this paper presents innovative research on BCI for speech communication. Since imagined speech suffers from quite a few factors, we have chosen to focus on subvocalized speech for the current work. The current work is considered to be the first to utilize the subvocal verbalization for EEG-based BCI in speech communication. The electrical signals generated by the human brain during subvocalized speech are captured, analyzed, and interpreted as speech. Further, the processed EEG signals are used to drive a speech synthesizer, enabling communication and acoustical feedback for the user. We attempt to demonstrate and justify that the BCI is capable of providing good results. The basis of this effort is the presumption that, whether the speech is overt or covert, it always originates in the mind. The scalp maps provide evidence that subvocal speech prediction, from the neurological signals, is achievable. The statistical results obtained from the current study demonstrate that speech prediction is possible. EEG signals suffer from the curse of dimensionality due to the intrinsic biological and electromagnetic complexities. Therefore, in the current work, the subset selection method, using pairwise cross-correlation, is proposed to reduce the size of the data while minimizing loss of information. The prominent variances obtained from the SSM, based on principal representative features, were deployed to analyze multiclass EEG signals. A multiclass support vector machine is used for the classification of EEG signals of five subvocalized words extracted from scalp electrodes. Though the current work identifies many challenges, the promise of this technology is exhibited.