EEG-based Syllable-Level Voice Activity Detection

Speech brain-computer interface (BCI), as an ideal means to achieve direct communication between the brain and the outside world, has become a research area of great interest. This work studied syllable-level voice activity detection (VAD) based on electroencephalogram (EEG) signals to help identify the presence or absence of speech-related EEG activity. We utilized EEG signals from 10 participants performing auditory (listening to stimuli) and speech (pronouncing syllables) tasks to measure brain activity. Speech-Based VAD was employed to label the auditory stimuli and voice recordings, generating corresponding brain activity labels, which were then used to classify resting and active (listening or pronouncing) EEG states, respectively. The experimental results showed that the EEG-based VAD model achieved accuracies of 90.93% and 69.57% for the speech production and auditory speech tasks, respectively. The accuracies were lower in the cross-subject classification, with accuracies of 72.63% and 61.15% for the two tasks. Additionally, the experiment further compared the model's performance under different time window conditions, but no significant correlation was found between window length and classification accuracy. This study provided new insights into the application of EEG based speech decoding, particularly in future self-paced speech BCI applications.