脑电（Electroencephalogram, EEG）为言语想象脑机接口（Brain-Computer Interface, BCI）提供了核心解码基础，使言语障碍患者无须外周神经与肌肉通路即可实现人机交互，在失语症康复与隐蔽通信等领域具有重要潜力。本文系统综述了EEG言语想象BCI的编码范式与解码算法进展：在编码方面，言语想象的神经模式与言语执行高度重叠，主要涉及Broca区、Wernicke区等语言网络节点，想象内容涵盖音位、音节、字词和句子，范式设计常借助多模态刺激与重复任务以提升信号可分性。解码方面，传统两阶段方法（特征提取+分类）受限于特征工程依赖与跨被试泛化能力弱的问题；端到端深度学习通过直接处理原始EEG时空序列，实现特征与分类联合优化，提升了解码效率。新兴方法融合自然语言处理与预训练语言模型，利用语义上下文实现句子层级解码，为复杂语言重建提供新路径。当前挑战包括解码精度不足与跨被试泛化性差，未来需结合少样本学习、多模态融合，并推动EEG与大语言模型协同，以促进临床应用。

Electroencephalogram (EEG) provides the core decoding foundation for speech imagery brain-computer interfaces (BCI), enabling individuals with speech impairments to achieve human-computer interaction without relying on peripheral nerves and muscle pathways. It holds significant potential in areas such as aphasia rehabilitation and covert communication. This article systematically reviews advances in neural encoding paradigms and decoding algorithms for EEG-based speech imagery BCI: In terms of encoding, the neural patterns during speech imagery highly overlap with those during speech execution, primarily involving key nodes of the language network such as Broca's area and Wernicke's area. The imagined content spans phonemes, syllables, words, and sentences. Paradigm design often employs multimodal stimuli and repetitive tasks to enhance signal separability. Regarding decoding, traditional two-stage methods (feature extraction classification) are limited by their dependence on feature engineering and poor cross-subject generalization. End-to-end deep learning models directly process raw spatiotemporal EEG sequences, achieving joint optimization of feature extraction and classification, which significantly improves decoding efficiency. Emerging approaches integrate natural language processing
and pre-trained language models, leveraging semantic context to achieve sentence-level decoding and offering new pathways for complex language reconstruction. Current key challenges include insufficient decoding accuracy and poor cross-subject generalization. Future efforts should incorporate few-shot learning frameworks to enhance model generalization, integrate
multimodal data to deepen the understanding of neural mechanisms underlying EEG, and promote synergy between EEG signals and large language models to advance practical clinical applications.