An open-access EEG dataset for speech decoding: Exploring the role of articulation and coarticulation

Area of Science:

• Neuroscience
• Linguistics
• Computer Science

Background:

• Electroencephalography (EEG) is a non-invasive technique for measuring neural activity, showing potential for brain-computer interface (BCI) applications.
• There is a growing need for standardized, publicly available datasets that capture the complexity of naturalistic speech for EEG-based BCI research.
• Existing EEG datasets often lack the linguistic complexity required for real-world BCI applications, and solutions must address signal noise and cross-session reliability.

Purpose of the Study:

• To present two validated electroencephalography (EEG) datasets for the development and evaluation of brain-computer interface (BCI) systems.
• To enable classification of neural signals at the phoneme and word levels, and by the articulatory properties of phonemes.
• To investigate the effect of transcranial magnetic stimulation (TMS) on EEG signals during speech processing.

Main Methods:

• Recorded 64-channel EEG data from 24 subjects (N=8 and N=16 in two datasets) while they listened to and repeated vowels and consonants.
• Generated stimuli including consonant-vowel pairs, real words, and pseudowords, incorporating coarticulation effects.
• Presented stimuli under control conditions and during transcranial magnetic stimulation (TMS) to assess its impact on EEG signals.

Main Results:

• Two validated EEG datasets were created, suitable for phoneme and word-level classification.
• The datasets include complex speech stimuli, such as consonant-vowel pairs and words, reflecting real-world linguistic variations.
• The study assessed the potential augmentation of EEG signals by TMS during articulatory processes.

Conclusions:

• The presented datasets provide a valuable resource for advancing EEG-based BCI research in naturalistic speech processing.
• These datasets will facilitate the development of more robust and reliable BCI systems capable of handling linguistic complexity.
• Further research can utilize these datasets to explore the neural correlates of speech production and perception, and the efficacy of TMS in BCI applications.