DCB‐Net: A hybrid depthwise CNN‐BiLSTM approach for fine‐grained imagined speech decoding using EEG signals

Abstract Imagined speech decoding from phonemes remains challenging owing to low signal‐to‐noise ratios, pronounced inter‐trial variability, and subtle spatiotemporal dynamics embedded in neural signals. In this study, DCB‐Net, a hybrid Depthwise CNN‐BiLSTM network, is proposed to jointly model channel‐wise spatial characteristics and long‐range temporal dependencies inherent to imagined speech neural signals. Unlike conventional neural decoding methods, DCB‐Net explicitly decomposes channel‐wise temporal feature extraction using depthwise separable convolution and bidirectional temporal context modeling using BiLSTM. This results in a compact, computationally efficient model for fine‐grained imagined‐speech phoneme classification. The proposed model is evaluated on the SpeechBCI imagined‐speech dataset comprising 40 phoneme classes. To validate classification performance and neurological reliability, extensive experimental analyses were conducted. DCB‐Net achieved a classification accuracy of 72.41%, outperforming state‐of‐the‐art methods while exhibiting improved phoneme‐level separability and consistent neural representation patterns.