EEG-Based Imagined-Speech Decoding: A Review

Non-invasive neural speech interfaces aim to reconstruct intended words directly from brain activity, providing vital communication channels for individuals with severe dysarthria or locked-in syndrome. Among various modalities, electroencephalography (EEG) remains the most accessible and cost-effective option. Nevertheless, decoding imagined speech from EEG signals remains challenging due to their inherently low signal-to-noise ratio, pronounced inter-subject variability, and the limited size and heterogeneity of available corpora. This review synthesises peer-reviewed studies on EEG-based imagined-speech recognition, encompassing preprocessing pipelines, feature extraction strategies, and deep learning-based decoding methods. Unlike earlier surveys that emphasised isolated algorithmic aspects, this review offers a comprehensive perspective across the entire decoding pipeline from neural acquisition and signal conditioning to representation learning and evaluation. Persistent gaps include inconsistent evaluation protocols, restricted vocabularies, and limited generalisation across subjects. Furthermore, the review summarises key findings and proposes a set of practical recommendations to guide reproducible progress in the field. Finally, it highlights the principal technical challenges of deep-learning-based imagined-speech decoding and outlines promising future research directions.

无创神经语音接口旨在直接从大脑活动中重构用户意欲表达的词汇，为患有严重构音障碍或闭锁综合征的群体提供至关重要的沟通渠道。在各类采集模态中，脑电图（electroencephalography, EEG）仍是最易获取且性价比最高的选择。然而，由于脑电信号本身信噪比极低、个体间差异显著，且现有语料库规模有限且异质性较强，从脑电信号中解码想象语音仍面临诸多挑战。本综述整合了基于脑电图的想象语音识别领域的同行评议研究，涵盖预处理流程、特征提取策略以及基于深度学习的解码方法。与此前仅聚焦单一算法层面的综述不同，本综述涵盖了从神经信号采集、信号调理到表征学习与模型评估的完整解码流程，提供了全面的研究视角。该领域仍存在诸多持续存在的研究空白，包括评估协议不统一、词汇表受限以及跨受试者泛化能力不足等问题。此外，本综述总结了该领域的核心研究成果，并提出了一系列实用建议，以推动该领域可复现的研究进展。最后，本综述着重指出了基于深度学习的想象语音解码所面临的主要技术挑战，并梳理了极具前景的未来研究方向。