Inner speech classification based on electroencephalography (EEG) signals and support vector machine (SVM)

Abstract Inner speech is a form of self-directed dialogue which plays an important role in cognitive development, speech monitoring, executive function, and psychopathology. Despite of a growing knowledge on its phenomenology, development, and function, approaches to the scientific study of inner speech have remained diffuse and largely unintegrated. Electroencephalography (EEG) which is a non-intrusive approach for brain-computer interface (BCI) brings new options to inner speech studies. Due to the advantages of EEG, more and more studies are related to inner speech apply EEG signals. In this contribution, different words expressed in inner speech are distinguished by applying EEG signals and support vector machine (SVM). Electroencephalography data from ‘Thinking out loud’ dataset open to public are employed. In the experiment, numerous EEG data are acquired from the 128 sensors located in the headcap. Therefore, data are filtered in the first step. Afterwards, selected data are decomposed by empirical mode decomposition (EMD) into various intrinsic mode functions (IMFs). Furthermore, IMFs are transformed using Hilbert transform to check the brain wave bands suitable for distinguishing inner speech. Lastly, single or combined of IMFs are classified by support vector machine (SVM) with various kernels. When most suitable IMFs and kernels are employed, the average results for each subject scheme are: F-score: 99.24 %, accuracy: 99.24 %, and standard deviation (SD): 0.95. Best results for all subject schemes are: F-score: 99.67 %, accuracy: 99.66 %, and standard deviation (SD): 0.27. The obtained results demonstrate that the proposed approach can differentiate EEG signals from inner speech well.