Synthesizing intelligible utterances from EEG of imagined speech

Decoding natural language directly from neural activity is of great interest to people with limited communication means. Being a non-invasive and convenient approach, the electroencephalogram (EEG) offers better portability and wider application potentiality. We present an EEG-to-speech system (ETS) that synthesizes audible, and highly understandable language by EEG of imagined speech. Our ETS applies a specially designed X-shape deep neural network (DNN) to realize an End-to-End correspondence between imagined speech EEG and the speech. The system innovatively incorporates dynamic time warping into the network's training process, using actual speech EEG data as a bridge to temporally align imagined speech EEG signals with speech signals. The ETS performance was evaluated on 13 participants who pretraining four Chinese disyllabic words. These words cover all four tones and 40% of the phonemes in Chinese. Our synthesized utterances' average accuracy across all subjects was 91.23%, the average MOS value was 3.50, and the best accuracy was 99% with an MOS value of 3.99. Furthermore, a partial feedback mechanism for DNN and spectral subtraction-based speech enhancement are introduced to improve the quality of generated speech. Our findings prove that non-invasive approaches can be a significant step in regaining verbal language ability.